December 4, 2014 by dominiccummings

Times op-ed: What Is To Be Done? An answer to Dean Acheson’s famous quip

On Tuesday 2 December, the Times ran an op-ed by me you can see HERE. It got cut slightly for space. Below is the original version that makes a few other points.

I will use this as a start of a new series on what can be done to improve the system including policy, institutions, and management.

NB1. The article is not about the election or party politics. My suggested answer to Acheson is, I think, powerful partly because it is something that could be agreed upon, in various dimensions, across the political spectrum. I left the DfE in January partly because I wanted to have nothing to do with the election and this piece should not be seen as advocating ‘something Tories should say for the election’. I do not think any of the three leaders are interested in or could usefully pursue this goal – I am suggesting something for the future when they are all gone, and they could quite easily all be gone by summer 2016.

NB2. My view is not – ‘public bad, private good’. As I explained in The Hollow Men II, a much more accurate and interesting distinction is between a) large elements of state bureaucracies, dreadful NGOs like the CBI, and many large companies (that have many of the same HR and incentive problems as bureaucracies), where very similar types rise to power because the incentives encourage political skills rather than problem-solving skills, and b) start-ups, where entrepreneurs and technically trained problem-solvers can create organisations that operate extremely differently, move extremely fast, create huge value, and so on.

(For a great insight into start-up world I recommend two books. 1. Peter Thiel’s new book ‘Zero To One‘. 2. An older book telling the story of a mid-90s start-up that was embroiled in the Netscape/Microsoft battle and ended up selling itself to the much better organised Bill Gates – ‘High Stakes, No Prisoners‘ by Charles Ferguson. This blog, Creators and Rulers, by physicist Steve Hsu also summarises some crucial issues excellently.)

Some parts of government can work like start-ups but the rest of the system tries to smother them. For example, DARPA (originally ARPA) was set up as part of the US panic about Sputnik. It operates on very different principles from the rest of the Pentagon’s R&D system. Because it is organised differently, it has repeatedly produced revolutionary breakthroughs (e.g. the internet) despite a relatively tiny budget. But also note – DARPA has been around for decades and its operating principles are clear but nobody else has managed to create an equivalent (openly at least). Also note that despite its track record, D.C. vultures constantly circle trying to make it conform to the normal rules or otherwise clip its wings. (Another interesting case study would be the alternative paths taken by a) the US government developing computers with one genius mathematician, von Neumann, post-1945 (a lot of ‘start-up’ culture) and b) the UK government’s awful decisions in the same field with another genius mathematician, Turing, post-1945.)

When I talk about new and different institutions below, this is one of the things I mean. I will write a separate blog just on DARPA but I think there are two clear action points:

1. We should create a civilian version of DARPA aimed at high-risk/high-impact breakthroughs in areas like energy science and other fundamental areas such as quantum information and computing that clearly have world-changing potential. For it to work, it would have to operate outside all existing Whitehall HR rules, EU procurement rules and so on – otherwise it would be as dysfunctional as the rest of the system (defence procurement is in a much worse state than the DfE, hence, for example, billions spent on aircraft carriers that in classified war-games cannot be deployed to warzones). We could easily afford this if we could prioritise – UK politicians spend far more than DARPA’s budget on gimmicks every year – and it would provide huge value with cascading effects through universities and businesses.

2. The lessons of why and how it works – such as incentivising goals, not micromanaging methods – have general application that are useful when we think generally about Whitehall reform.

Finally, government institutions also operate to exclude from power scientists, mathematicians, and people from the start-up world – the Creators, in Hsu’s term. We need to think very hard about how to use their very rare and valuable skills as a counterweight to the inevitable psychological type that politics will always tend to promote.

Please leave comments, corrections etc below.

What Is to Be Done?

There is growing and justified contempt for Westminster. Number Ten has become a tragi-comic press office with the prime minister acting as Über Pundit. Cameron, Miliband, and Clegg see only the news’s flickering shadows on their cave wall – they cannot see the real world behind them. As they watch floundering MPs, officials know they will stay in charge regardless of an election that won’t significantly change Britain’s trajectory.

Our institutions failed pre-1914, pre-1939, and with Europe. They are now failing to deal with a combination of debts, bad public services, security threats, and profound transitions in geopolitics, economics, and technology. They fail in crises because they are programmed to fail. The public knows we need to reorient national policy and reform these institutions. How?

First, we need a new goal. In 1962, Dean Acheson quipped that Britain had failed to find a post-imperial role. The romantic pursuit of ‘the special relationship’ and the deluded pursuit of a leading EU role have failed. This role should focus on making Britain the best country for education and science. Pericles described Athens as ‘the school of Greece’: we could be the school of the world because this role depends on thought and organisation, not size.

This would give us a central role in tackling humanity’s biggest problems and shaping the new institutions, displacing the EU and UN, that will emerge as the world makes painful transitions in coming decades. It would provide a focus for financial priorities and Whitehall’s urgent organisational surgery. It’s a goal that could mobilise very large efforts across political divisions as the pursuit of knowledge is an extremely powerful motive.

Second, we must train aspirant leaders very differently so they have basic quantitative skills and experience of managing complex projects. We should stop selecting leaders from a subset of Oxbridge egomaniacs with a humanities degree and a spell as spin doctor.

In 2012, Fields Medallist Tim Gowers sketched a ‘maths for presidents’ course to teach 16-18 year-olds crucial maths skills, including probability and statistics, that can help solve real problems. It starts next year. [NB. The DfE funded MEI to turn this blog into a real course.] A version should be developed for MPs and officials. (A similar ‘Physics for Presidents‘ course has been a smash hit at Berkeley.) Similarly, pioneering work by Philip Tetlock on ‘The Good Judgement Project‘ has shown that training can reduce common cognitive errors and can sharply improve the quality of political predictions, hitherto characterised by great self-confidence and constant failure.

New interdisciplinary degrees such as ‘World history and maths for presidents’ would improve on PPE but theory isn’t enough. If we want leaders to make good decisions amid huge complexity, and learn how to build great teams, then we should send them to learn from people who’ve proved they can do it. Instead of long summer holidays, embed aspirant leaders with Larry Page or James Dyson so they can experience successful leadership.

Third, because better training can only do so much, we must open political institutions to people and ideas from outside SW1.

A few people prove able repeatedly to solve hard problems in theoretical and practical fields, creating important new ideas and huge value. Whitehall and Westminster operate to exclude them from influence. Instead, they tend to promote hacks and apparatchiks and incentivise psychopathic narcissism and bureaucratic infighting skills – not the pursuit of the public interest.

How to open up the system? First, a Prime Minister should be able to appoint Secretaries of State from outside Parliament. [How? A quick and dirty solution would be: a) shove them in the Lords, b) give Lords ministers ‘rights of audience’ in the Commons, c) strengthen the Select Committee system.]

Second, the 150 year experiment with a permanent civil service should end and Whitehall must open to outsiders. The role of Permanent Secretary should go and ministers should appoint departmental chief executives so they are really responsible for policy and implementation. Expertise should be brought in as needed with no restrictions from the destructive civil service ‘human resources’ system that programmes government to fail. Mass collaborations are revolutionising science [cf. Michael Nielsen’s brilliant book]; they could revolutionise policy. Real openness would bring urgent focus to Whitehall’s disastrous lack of skills in basic functions such as budgeting, contracts, procurement, legal advice, and project management.

Third, Whitehall’s functions should be amputated. The Department for Education improved as Gove shrank it. Other departments would benefit from extreme focus, simplification, and firing thousands of overpaid people. If the bureaucracy ceases to be ‘permanent’, it can adapt quickly. Instead of obsessing on process, distorting targets, and micromanaging methods, it could shift to incentivising goals and decentralising methods.

Fourth, existing legal relationships with the EU and ECHR must change. They are incompatible with democratic and effective government

Fifth, Number Ten must be reoriented from ‘government by punditry’ to a focus on the operational planning and project management needed to convert priorities to reality over months and years.

Technological changes such as genetic engineering and machine intelligence are bringing revolution. It would be better to undertake it than undergo it.

September 10, 2014 by dominiccummings

Complexity, ‘fog and moonlight’, prediction, and politics II: controlled skids and immune systems (UPDATED)

‘Politics is a job that can really only be compared with navigation in uncharted waters. One has no idea how the weather or the currents will be or what storms one is in for. In politics, there is the added fact that one is largely dependent on the decisions of others, decisions on which one was counting and which then do not materialise; one’s actions are never completely one’s own. And if the friends on whose support one is relying change their minds, which is something that one cannot vouch for, the whole plan miscarries… One’s enemies one can count on – but one’s friends!’ Otto von Bismarck.

‘Everything in war is very simple, but the simplest thing is difficult. The difficulties accumulate and end by producing a kind of friction that is inconceivable unless one has experienced war… Countless minor incidents – the kind you can never really foresee – combine to lower the general level of performance, so that one always falls short of the intended goal. Iron will-power can overcome this friction … but of course it wears down the machine as well… Friction is the only concept that … corresponds to the factors that distinguish real war from war on paper. The … army and everything else related to it is basically very simple and therefore seems easy to manage. But … each part is composed of individuals, every one of whom retains his potential of friction… This tremendous friction … is everywhere in contact with chance, and brings about effects that cannot be measured… Friction … is the force that makes the apparently easy so difficult… Finally … all action takes place … in a kind of twilight, which like fog or moonlight, often tends to make things seem grotesque and larger than they really are. Whatever is hidden from full view in this feeble light has to be guessed at by talent, or simply left to chance.’ Clausewitz.

In July, I wrote a blog on complexity and prediction which you can read HERE.

I will summarise briefly its main propositions and add some others. All page references are to my essay, HERE. (Section 1 explores some of the maths and science issues below in more detail.)

Some people asked me after Part I – why is such abstract stuff important to practical politics? That is a big question but in a nutshell…

If you want to avoid the usual fate in politics of failure, you need to understand some basic principles about why people make mistakes and how some people, institutions, and systems cope with mistakes and thereby perform much better than most. The reason why Whitehall is full of people failing in predictable ways on an hourly basis is because, first, there is general system-wide failure and, second, everybody keeps their heads down focused on the particular and they ignore the system. Officials who speak out see their careers blow up. MPs are so cowed by the institutions and the scale of official failure that they generally just muddle along tinkering and hope to stay a step ahead of the media. Some understand the epic scale of institutional failure but they know that the real internal wiring of the system in the Cabinet Office has such a tight grip that significant improvement will be very hard without a combination of a) a personnel purge and b) a fundamental rewiring of power at the apex of the state. Many people in Westminster are now considering how this might happen. Such thoughts must, I think, be based on some general principles otherwise they are likely to miss the real causes of system failure and what to do.

In future blogs in this series, I will explore some aspects of markets and science that throw light on the question: how can humans and their institutions cope with these problems of complexity, uncertainty, and prediction in order to limit failures?

Separately, The Hollow Men II will focus on specifics of how Whitehall and Westminster work, including Number Ten and some examples from the Department for Education.

Considering the more general questions of complexity and prediction sheds light on why government is failing so badly and how it could be improved.

Complexity, nonlinearity, uncertainty, and prediction

Even the simplest practical problems are often very complex. If a Prime Minister wants to line up 70 colleagues in Downing Street to blame them for his woes, there are 70! ways of lining them up and 70! [70! = 70 x 69 x 68 … x 2 x 1] is roughly 10¹⁰⁰ (a ‘googol’), which is roughly ten billion times the estimated number of atoms in the universe (10⁹⁰). [See comments.]

Even the simplest practical problems, therefore, can be so complicated that searching through the vast landscape of all possible solutions is not practical.

After Newton, many hoped that perfect prediction would be possible:

‘An intellect which at a certain moment would know all the forces that animate nature, and all positions of the beings that compose it, if this intellect were vast enough to submit the data to analysis, would condense in a single formula the movements of the greatest bodies of the universe and those of the tiniest atom; for such an intellect nothing would be uncertain and the future just like the past would be present before its eyes’ (Laplace).

However, most of the most interesting systems in the world – such as brains, cultures, and conflicts – are nonlinear. That is, a small change in input has an arbitrarily large affect on output. Have you ever driven through a controlled skid then lost it? A nonlinear system is one in which you can shift from ‘it feels great on the edge’ to ‘I’m steering into the skid but I’ve lost it and might die in a few seconds’ because of one tiny input change, like your tyre catches a cat’s eye in the wet. This causes further problems for prediction. Not only is the search space so vast it cannot be searched exhaustively, however fast our computers, but in nonlinear systems one has the added problem that a tiny input change can lead to huge output changes.

Some nonlinear systems are such that no possible accuracy of measurement of the current state can eliminate this problem – there is unavoidable uncertainty about the future state. As Poincaré wrote, ‘it may happen that small differences in the initial conditions produce very great ones in the final phenomena. A small error in the former will produce an enormous error in the latter. Prediction becomes impossible, and we have the fortuitous phenomenon.’ It does not matter that the measurement error is in the 20^th decimal place – the prediction will still quickly collapse.

Weather systems are like this which is why, despite the enormous progress made with predictions, we remain limited to ~10-14 days at best. To push the horizon forward by just one day requires exponential increases in the resources required. Political systems are also nonlinear. If Cohen-Blind’s aim had been very slightly different in May 1866 when he fired five bullets at Bismarck, the German states would certainly have evolved in a different way and perhaps there would have been no fearsome German army led by a General Staff into World War I, no Lenin and Hitler, and so on. Bismarck himself appreciated this very well. ‘We are poised on the tip of a lightning conductor, and if we lose the balance I have been at pains to create we shall find ourselves on the ground,’ he wrote to his wife during the 1871 peace negotiations in Versailles. Social systems are also nonlinear. Online experiments have explored how complex social networks cannot be predicted because of initial randomness combining with the interdependence of decisions.

In short, although we understand some systems well enough to make precise or statistical predictions, most interesting systems – whether physical, mental, cultural, or virtual – are complex, nonlinear, and have properties that emerge from feedback between many interactions. Exhaustive searches of all possibilities are impossible. Unfathomable and unintended consequences dominate. Problems cascade. Complex systems are hard to understand, predict and control.

Humans evolved in this complex environment amid the sometimes violent, sometimes cooperative sexual politics of small in-groups competing with usually hostile out-groups. We evolved to sense information, process it, and act. We had to make predictions amid uncertainty and update these predictions in response to feedback from our environment – we had to adapt because we have necessarily imperfect data and at best approximate models of reality. It is no coincidence that in one of the most famous speeches in history, Pericles singled out the Athenian quality of adaptation (literally ‘well-turning’) as central to its extraordinary cultural, political and economic success.

How do we make these predictions, how do we adapt? Much of how we operate depends on relatively crude evolved heuristics (rules of thumb) such as ‘sense movement >> run/freeze’. These heuristics can help. Further, our evolved nature gives us amazing pattern recognition and problem-solving abilities. However, some heuristics lead to errors, illusions, self-deception, groupthink and so on – problems that often swamp our reasoning and lead to failure.

I will look briefly at a) the success of science and mathematical models, b) the success of decentralised coordination in nature and markets, and c) the failures of political prediction and decision-making.

The success of science and mathematical models

Our brains evolved to solve social and practical problems, not to solve mathematical problems. This is why translating mathematical and logical problems into social problems makes them easier for people to solve (cf. Nielsen.) Nevertheless, a byproduct of our evolution was the ability to develop maths and science. Maths gives us an abstract structure of certain knowledge that we can use to build models of the world. ‘[S]ciences do not try to explain, they hardly even try to interpret, they mainly make models. By a model is meant a mathematical construct which, with the addition of certain verbal interpretations, describes observed phenomena. The justification of such a mathematical construct is solely and precisely that it is expected … correctly to describe phenomena from a reasonably wide area’ (von Neumann).

Because the universe operates according to principles that can be approximated by these models, we can understand it approximately. ‘Why’ is a mystery. Why should ‘imaginary numbers’ based on the square root of minus 1, conceived five hundred years ago and living for hundreds of years without practical application, suddenly turn out to be necessary in the 1920s to calculate how subatomic particles behave? How could it be that in a serendipitous meeting in the IAS cafeteria in 1972, Dyson and Montgomery should realise that an equation describing the distribution of prime numbers should also describe the energy level of particles? We can see that the universe displays a lot of symmetry but we do not know why there is some connection between the universe’s operating principles and our evolved brains’ abilities to do abstract mathematics. Einstein asked, ‘How is it possible that mathematics, a product of human thought that is independent of experience, fits so excellently the objects of physical reality?’ Wigner replied to Einstein in a famous paper, ‘The Unreasonable Effectiveness of Mathematics in the Natural Sciences’ (1960) but we do not know the answer. (See ‘Is mathematics invented or discovered?’, Tim Gowers, 2011.)

The accuracy of many of our models gets better and better. In some areas such as quantum physics, the equations have been checked so delicately that, as Feynman said, ‘If you were to measure the distance from Los Angeles to New York to this accuracy, it would be exact to the thickness of a human hair’. In other areas, we have to be satisfied with statistical models. For example, many natural phenomenon, such as height and intelligence, can be modelled using ‘normal distributions’. Other phenomena, such as the network structure of cells, the web, or banks in an economy, can be modelled using ‘power laws’. [* See End] Why do statistical models work? Because ‘chance phenomena, considered collectively and on a grand scale, create a non-random regularity’ (Kolmogorov). [** See End]

Science has also built an architecture for its processes, involving meta-rules, that help correct errors and normal human failings. For example, after Newton the system of open publishing and peer review developed. This encouraged scientists to make their knowledge public, confident that they would get credit (instead of hiding things in code like Newton). Experiments must be replicated and scientists are expected to provide their data honestly so that others can test their claims, however famous, prestigious, or powerful they are. Feynman described the process in physics as involving, at its best, ‘a kind of utter honesty … [Y]ou should report everything that you think might make [your experiment or idea] invalid… [Y]ou must also put down all the facts which disagree with it, as well as those that agree with it… The easiest way to explain this idea is to contrast it … with advertising.’

The architecture of the scientific process is not perfect. Example 1. Evaluation of contributions is hard. The physicist who invented the arXiv was sacked soon afterwards because his university’s tick box evaluation system did not have a way to value his enormous contribution. Example 2. Supposedly ‘scientific’ advice to politicians can also be very overconfident. E.g. A meta-study of 63 studies of the costs of various energy technologies reveals: ‘The discrepancies between equally authoritative, peer-reviewed studies span many orders of magnitude, and the overlapping uncertainty ranges can support almost any ranking order of technologies, justifying almost any policy decision as science based’ (Stirling, Nature, 12/2010).

This architecture and its meta-rules are now going through profound changes, brilliantly described by the author of the seminal textbook on quantum computers, Michael Nielsen, in his book Reinventing Discovery – a book that has many lessons for the future of politics too. But overall the system clearly has great advantages.

The success of decentralised information processing in solving complex problems

Complex systems and emergent properties

Many of our most interesting problems can be considered as networks. Individual nodes (atoms, molecules, genes, cells, neurons, minds, organisms, organisations, computer agents) and links (biochemical signals, synapses, internet routers, trade routes) form physical, mental, and cultural networks (molecules, cells, organisms, immune systems, minds, organisations, internet, biosphere, ‘econosphere’, cultures) at different scales.

The most interesting networks involve interdependencies (feedback and feedforward) – such as chemical signals, a price collapse, neuronal firing, an infected person gets on a plane, or an assassination – and are nonlinear. Complex networks have emergent properties including self-organisation. For example, the relative strength of a knight in the centre of the chessboard is not specified in the rules but emerges from the nodes of the network (or ‘agents’) operating according to the rules.

Even in physics, ‘The behavior of large and complex aggregates of elementary particles … is not to be understood in terms of a simple extrapolation of the properties of a few particles. Instead, at each level of complexity entirely new properties appear’ (Anderson). This is more obvious in biological and social networks.

Ant colonies and immune systems: how decentralised information processing solves complex problems

Ant colonies and the immune system are good examples of complex nonlinear systems with ‘emergent properties’ and self-organisation.

The body cannot ‘know’ in advance all the threats it will face so the immune system cannot be perfectly ‘pre-designed’. How does it solve this problem?

There is a large diverse population of individual white blood cells (millions produced per day) that sense threats. If certain cells detect that a threat has passed a threshold, then they produce large numbers of daughter cells, with mutations, that are tested on captured ‘enemy’ cells. Unsuccessful daughter cells die while successful ones are despatched to fight. These daughter cells repeat the process so a rapid evolutionary process selects and reproduces the best defenders and continually improves performance. Other specialist cells roam around looking for invaders that have been tagged by antibodies. Some of the cells remain in the bloodstream, storing information about the attack, to guard against future attacks (immunity).

There is a constant evolutionary arms race against bacteria and other invaders. Bacteria take over cells’ machinery and communications. They reprogram cells to take them over or trigger self-destruction. They disable immune cells and ‘ride’ them back into lymph nodes (Trojan horse style) where they attack. They shape-change fast so that immune cells cannot recognise them. They reprogram immune cells to commit suicide. They reduce competition by tricking immune cells into destroying other bacteria that help the body fight infection (e.g. by causing diarrhoea to flush out competition).

NB. there is no ‘plan’ and no ‘central coordination’. The system experiments probabilistically, reinforces success, and discards failure. It is messy. Such a system cannot be based on trying to ‘eliminate failure’. It is based on accepting a certain amount of failure but keeping it within certain tolerances via learning.

Looking at an individual ant, it would be hard to know that an ant colony is capable of farming, slavery, and war.

‘The activity of an ant colony is totally defined by the activities and interactions of its constituent ants. Yet the colony exhibits a flexibility that goes far beyond the capabilities of its individual constituents. It is aware of and reacts to food, enemies, floods, and many other phenomena, over a large area; it reaches out over long distances to modify its surroundings in ways that benefit the colony; and it has a life-span orders of magnitude longer than that of its constituents… To understand the ant, we must understand how this persistent, adaptive organization emerges from the interactions of its numerous constituents.’ (Hofstadter)

Ant colonies face a similar problem to the immune system: they have to forage for food in an unknown environment with an effectively infinite number of possible ways to search for a solution. They send out agents looking for food; those that succeed return to the colony leaving a pheromone trail which is picked up by others and this trail strengthens. Decentralised decisions via interchange of chemical signals drive job-allocation (the division of labour) in the colony. Individual ants respond to the rate of what others are doing: if an ant finds a lot of foragers, it is more likely to start foraging.

Similarities between the immune system and ant colonies in solving complex problems

Individual white blood cells cannot access the whole picture; they sample their environment via their receptors. Individual ants cannot cannot access the whole picture; they sample their environment via their chemical processors. The molecular shape of immune cells and the chemical processing abilities of ants are affected by random mutations; the way individual cells or ants respond has a random element. The individual elements (cells / ants) are programmed to respond probabilistically to new information based on the strength of signals they receive.

Environmental exploration by many individual agents coordinated via feedback signals allows a system to probe many different probabilities, reinforce success, ‘learn’ from failure (e.g withdraw resources from unproductive strategies), and keep innovating (e.g novel cells are produced even amid a battle and ants continue to look for better options even after striking gold). ‘Redundancy’ allows local failures without breaking the system. There is a balance between exploring the immediate environment for information and exploiting that information to adapt.

In such complex networks with emergent properties, unintended consequences dominate. Effects cascade: ‘they come not single spies but in battalions’. Systems defined as ‘tightly coupled‘ – that is, they have strong interdependencies so that the behaviour of one element is closely connected to another – are not resilient in the face of nonlinear events (picture a gust of wind knocking over one domino in a chain).

Network topology

We are learning how network topology affects these dynamics. Many networks (including cells, brains, the internet, the economy) have a topology such that nodes are distributed according to a power law (not a bell curve), which means that the network looks like a set of hubs and spokes with a few spokes connecting hubs. This network topology makes them resilient to random failure but vulnerable to the failure of critical hubs that can cause destructive cascades (such as financial crises) – an example of the problems that come with nonlinearity.

Similar topology and dynamics can be seen in networks operating at very different scales ranging from cellular networks, the brain, the financial system, the economy in general, and the internet. Disease networks often shows the same topology, with certain patients, such as those who get on a plane from West Africa to Europe with Ebola, playing the role of critical hubs connecting different parts of the network. Terrorist networks also show the same topology. All of these complex systems with emergent properties have the same network topology and are vulnerable to the failure of critical hubs.

Many networks evolve modularity. A modular system is one in which specific modules perform specific tasks, with links between them allowing broader coordination. This provides greater effectiveness and resilience to shocks. For example, Chongqing in China saw the evolution of a new ecosystem for designing and building motorbikes in which ‘assembler’ companies assemble modular parts built by competing companies, instead of relying on high quality vertically integrated companies like Yamaha. This rapidly decimated Japanese competition. Connections between network topology, power laws and fractals can be seen in work by physicist Geoffrey West both on biology and cities, for it is clear that just as statistical tools like the Central Limit Theorem demonstrate similar structure in completely different systems and scales, so similar processes occur in biology and social systems. [See Endnote.]

Markets: how decentralised information processing solves complex problems

[Coming imminently]

A summary of the progress brought by science and markets

The combination of reasoning, reliable accumulated knowledge, and a reliable institutional architecture brings steady progress, and occasional huge breakthroughs and wrong turns, in maths and science. The combination of the power of decentralised information processing to find solutions to complex problems and an institutional architecture brings steady progress, and occasional huge breakthroughs and wrong turns, in various fields that operate via markets.

Fundamental to the institutional architecture of markets and science is mechanisms that enable adaptation to errors. The self-delusion and groupthink that is normal for humans – being a side-effect of our nature as evolved beings – is partly countered by tried and tested mechanisms. These mechanisms are not based on an assumption that we can ‘eliminate failure’ (as so many in politics absurdly claim they will do). Instead, the assumption is that failure is a persistent phenomenon in a complex nonlinear world and it must be learned from and adapted to as quickly as possible. Entrepreneurs and scientists can be vain, go mad, or be prone to psychopathy – like public servants – but we usually catch it quicker and it causes less trouble. Catching errors, we inch forward ‘standing on the shoulders of giants’ as Newton put it.

Science has enabled humans to make transitions from numerology to mathematics, from astrology to astronomy, from alchemy to chemistry, from witchcraft to neuroscience, from tallies to quantum computation. Markets have been central to a partial transition in a growing fraction of the world from a) small, relatively simple, hierarchical, primitive, zero-sum hunter-gatherer tribes based on superstition (almost total ignorance of complex systems), shared aims, personal exchange and widespread violence, to b) large, relatively complex, decentralised, technological, nonzero-sum market-based cultures based on science (increasingly accurate predictions and control in some fields), diverse aims, impersonal exchange, trade, private property, and (roughly) equal protection under the law.

The failures of politics: wrong predictions, no reliable mechanisms for fixing obvious errors

‘No official estimates even mentioned that the collapse of Communism was a distinct possibility until the coup of 1989.’ National Security Agency, ‘Dealing With the Future’, declassified report.

However, the vast progress made in so many fields is clearly not matched in standards of government. In particular, it is very rare for individuals or institutions to make reliable predictions.

The failure of prediction in politics

Those in leading positions in politics and public service have to make all sorts of predictions. Faced with such complexity, politicians and others have operated mostly on heuristics (‘political philosophy’), guesswork, willpower and tactical adaptation. My own heuristics for working in politics are: focus, ‘know yourself’ (don’t fool yourself), think operationally, work extremely hard, don’t stick to the rules, and ask yourself ‘to be or to do?’.

Partly because politics is a competitive enterprise in which explicit and implicit predictions elicit countermeasures, predictions are particularly hard. This JASON report (PDF) on the prediction of rare events explains some of the technical arguments about predicting complex nonlinear systems such as disasters. Unsurprisingly, so-called ‘political experts’ are not only bad at predictions but are far worse than they realise. There are many prominent examples. Before the 2000 election, the American Political Science Association’s members unanimously predicted a Gore victory. Beyond such examples, we have reliable general data on this problem thanks to a remarkable study by Philip Tetlock. He charted political predictions made by supposed ‘experts’ (e.g will the Soviet Union collapse, will the euro collapse) for fifteen years from 1987 and published them in 2005 (‘Expert Political Judgement’). He found that overall, ‘expert’ predictions were about as accurate as monkeys throwing darts at a board. Experts were very overconfident: ~15 percent of events that experts claimed had no chance of occurring did happen, and ~25 percent of those that they said they were sure would happen did not happen. Further, the more media interviews an expert did, the less likely they were to be right. Specific expertise in a particular field was generally of no value; experts on Canada were about as accurate on the Soviet Union as experts on the Soviet Union were.

However, some did better than others. He identified two broad categories of predictor. The first he called ‘hedgehogs’ – fans of Big Ideas like Marxism, less likely to admit errors. The second he called ‘foxes’ – not fans of Big Ideas, more likely to admit errors and change predictions because of new evidence. (‘The fox knows many little things, but the hedgehog knows one big thing,’ Archilochus.) Foxes tended to make better predictions. They are more self-critical, adaptable, cautious, empirical, and multidisciplinary. Hedgehogs get worse as they acquire more credentials while foxes get better with experience. The former distort facts to suit their theories; the latter adjust theories to account for new facts.

Tetlock believes that the media values characteristics (such as Big Ideas, aggressive confidence, tenacity in combat and so on) that are the opposite of those prized in science (updating in response to new data, admitting errors, tenacity in pursuing the truth and so on). This means that ‘hedgehog’ qualities are more in demand than ‘fox’ qualities, so the political/media market encourages qualities that make duff predictions more likely. ‘There are some academics who are quite content to be relatively anonymous. But there are other people who aspire to be public intellectuals, to be pretty bold and to attach non-negligible probabilities to fairly dramatic change. That’s much more likely to bring you attention’ (Tetlock).

Tetlock’s book ought to be much-studied in Westminster particularly given 1) he has found reliable ways of identifying a small number of people who are very good forecasters and 2) IARPA (the intelligence community’s DARPA twin) is working with Tetlock to develop training programmes to improve forecasting skills. [See Section 6.] Tetolock says, ‘We now have a significant amount of evidence on this, and the evidence is that people can learn to become better. It’s a slow process. It requires a lot of hard work, but some of our forecasters have really risen to the challenge in a remarkable way and are generating forecasts that are far more accurate than I would have ever supposed possible from past research in this area.’ (This is part of IARPA’s ACE programme to develop aggregated forecast systems and crowdsourced prediction software. IARPA also has the SHARP programme to find ways to improve problem-solving skills for high-performing adults.)

His main advice? ‘If I had to bet on the best long-term predictor of good judgement among the observers in this book, it would be their commitment – their soul-searching Socratic commitment – to thinking about how they think’ (Tetlock). His new training programmes help people develop this ‘Socratic commitment’ and correct their mistakes in quite reliable ways.

NB. The extremely low quality of political forecasting is what allowed an outsider like Nate Silver to transform the field simply by applying some well-known basic maths.

The failure of prediction in economics

‘… the evidence from more than fifty years of research is conclusive: for a large majority of fund managers, the selection of stocks is more like rolling dice than like playing poker. Typically at least two out of every three mutual funds underperform the overall market in any given year. More important, the year-to-year correlation between the outcomes of mutual funds is very small, barely higher than zero. The successful funds in any given year are mostly lucky; they have a good roll the dice.’ Daniel Kahneman, winner of the economics ‘Nobel’ (not the same as the Nobel for physical sciences).

‘I importune students to read narrowly within economics, but widely in science…The economic literature is not the best place to find new inspiration beyond these traditional technical methods of modelling’ Vernon Smith, winner of the economics ‘Nobel’.

I will give a few examples of problems with economic forecasting.

In the 1961 edition of his famous standard textbook used by millions of students, one of the 20^th Century’s most respected economists, Paul Samuelson, predicted that respective growth rates in America and the Soviet Union meant the latter would overtake the USA between 1984-1997. By 1980, he had delayed the date to be in 2002-2012. Even in 1989, he wrote, ‘The Soviet economy is proof that, contrary to what many skeptics had earlier believed, a socialist command economy can function and even thrive.’

Chart: Samuelson’s prediction for the Soviet economy

The recent financial crisis also demonstrated many failed predictions. Various people, including physicists Steve Hsu and Eric Weinstein, published clear explanations of the extreme dangers in the financial markets and parallels with previous crashes such as Japan’s. However, they were almost totally ignored by politicians, officials, central banks and so on. Many of those involved were delusional. Perhaps most famously, Joe Cassano of AIG Financial said in a conference call (8/2007): ‘It’s hard for us – without being flippant – to even see a scenario within any kind of realm of reason that would see us losing one dollar in any of those transactions… We see no issues at all emerging.’

Nate Silver recently summarised some of the arguments over the crash and its aftermath. In December 2007, economists in the Wall Street Journal forecasting panel predicted only a 38 percent chance of recession in 2008. The Survey of Professional Forecasters is a survey of economists’ predictions done by the Federal Reserve Bank that includes uncertainty measurements. In November 2007, the Survey showed a net prediction by economists that the economy would grow by 2.4% in 2008, with a less than 3% chance of any recession and a 1-in-500 chance of it shrinking by more than 2%.

Chart: the 90% ‘prediction intervals’ for the Survey of Professional Forecasters net forecast of GDP growth 1993-2010

If the economists’ predictions were accurate, the 90% prediction interval should be right nine years out of ten, and 18 out of 20. Instead, the actual growth was outside the 90% prediction interval six times out of 18, often by a lot. (The record back to 1968 is worse.) The data would later reveal that the economy was already in recession in the last quarter of 2007 and, of course, the ‘1-in-500’ event of the economy shrinking by more than 2% is exactly what happened.**

Although the total volume of home sales in 2007 was only ~$2 trillion, Wall Street’s total volume of trades in mortgage-backed securities was ~$80 trillion because of the creation of ‘derivative’ financial instruments. Most people did not understand 1) how likely a house price fall was, 2) how risky mortgage-backed securities were, 3) how widespread leverage could turn a US housing crash into a major financial crash, and 4) how deep the effects of a major financial crash were likely to be. ‘The actual default rates for CDOs were more than two hundred times higher than S&P had predicted’ (Silver). In the name of ‘transparency’, S&P provided the issuers with copies of their ratings software allowing CDO issuers to experiment on how much junk they could add without losing a AAA rating. S&P even modelled a potential housing crash of 20% in 2005 and concluded its highly rated securities could ‘weather a housing downturn without suffering a credit rating downgrade.’

Unsurprisingly, Government unemployment forecasts were also wrong. Historically, the uncertainty in an unemployment rate forecast made during a recession had been about plus or minus 2 percent but Obama’s team, and economists in general, ignored this record and made much more specific predictions. In January 2009, Obama’s team argued for a large stimulus and said that, without it, unemployment, which had been 7.3% in December 2008, would peak at ~9% in early 2010, but with the stimulus it would never rise above 8% and would fall from summer 2009. However, the unemployment numbers after the stimulus was passed proved to be even worse than the ‘no stimulus’ prediction. Similarly, the UK Treasury’s forecasts about growth, debt, and unemployment from 2007 were horribly wrong but that has not stopped it making the same sort of forecasts.

Paul Krugman concluded from this episode: the stimulus was too small. Others concluded it had been a waste of money. Academic studies vary widely in predicting the ‘return’ from each $1 of stimulus. Since economists cannot even accurately predict a recession when the economy is already in recession, it seems unlikely that there will be academic consensus soon on such issues. Economics often seems like a sort of voodoo for those in power – spurious precision and delusions that there are sound mathematical foundations for the subject without a proper understanding of the conditions under which mathematics can help (cf. Von Neumann on maths and prediction in economics HERE).

Fields which do better at prediction

Daniel Kahneman, who has published some of the most important research about why humans make bad predictions, summarises the fundamental issues about when you can trust expert predictions:

‘To know whether you can trust a particular intuitive judgment, there are two questions you should ask: Is the environment in which the judgment is made sufficiently regular to enable predictions from the available evidence? The answer is yes for diagnosticians, no for stock pickers. Do the professionals have an adequate opportunity to learn the cues and the regularities? The answer here depends on the professionals’ experience and on the quality and speed with which they discover their mistakes. Anesthesiologists have a better chance to develop intuitions than radiologists do. Many of the professionals we encounter easily pass both tests, and their off-the-cuff judgments deserve to be taken seriously. In general, however, you should not take assertive and confident people at their own evaluation unless you have independent reason to believe that they know what they are talking about.’ (Emphasis added.)

It is obvious that politics fulfils neither of his two criteria – it does not even have hard data and clear criteria for success, like stock picking.

I will explore some of the fields that do well at prediction in a future blog.

The consequences of the failure of politicians and other senior decision-makers and their institutions

‘When superior intellect and a psychopathic temperament coalesce …, we have the best possible conditions for the kind of effective genius that gets into the biographical dictionaries’ (William James).

‘We’re lucky [the Unabomber] was a mathematician, not a molecular biologist’ (Bill Joy, Silicon Valley legend, author of ‘Why the future doesn’t need us’).

While our ancestor chiefs understood bows, horses, and agriculture, our contemporary chiefs (and those in the media responsible for scrutiny of decisions) generally do not understand their equivalents, and are often less experienced in managing complex organisations than their predecessors.

The consequences are increasingly dangerous as markets, science and technology disrupt all existing institutions and traditions, and enhance the dangerous potential of our evolved nature to inflict huge physical destruction and to manipulate the feelings and ideas of many people (including, sometimes particularly, the best educated) through ‘information operations’. Our fragile civilisation is vulnerable to large shocks and a continuation of traditional human politics as it was during 6 million years of hominid evolution – an attempt to secure in-group cohesion, prosperity and strength in order to dominate or destroy nearby out-groups in competition for scarce resources – could kill billions. We need big changes to schools, universities, and political and other institutions for their own sake and to help us limit harm done by those who pursue dreams of military glory, ‘that attractive rainbow that rises in showers of blood’ (Lincoln).

The global population of people with an IQ four standard deviations above the average (i.e. >160) is ~250k. About 1% of the population are psychopaths so there are perhaps ~2-3,000 with IQ ≈ Nobel/Fields winner. The psychopathic +3SD IQ (>145; average science PhD ~130) population is 30 times bigger. A subset will also be practically competent. Some of them may think, ‘Flectere si nequeo superos, / Acheronta movebo’ (‘If Heav’n thou can’st not bend, Hell thou shalt move’, the Aeneid). Board et al (2005) showed that high-level business executives are more likely than inmates of Broadmoor to have one of three personality disorders (PDs): histrionic PD, narcissistic PD, and obsessive-compulsive PD. Mullins-Sweatt et al (2010) showed that successful psychopaths are more conscientious than the unsuccessful.

A brilliant essay (here) by one of the 20th Century’s best mathematicians, John von Neumann, describes these issues connecting science, technology, and how institutions make decisions.

Some conclusions

When we consider why institutions are failing and how to improve them, we should consider the general issues discussed above. How to adapt quickly to new information? Does the institution’s structure incentivise effective adaptation or does it incentivise ‘fooling oneself’ and others? Is it possible to enable distributed information processing to find a ‘good enough’ solution in a vast search space? If your problem is similar to that of the immune system or ant colony, why are you trying to solve it with a centralised bureaucracy?

Further, some other obvious conclusions suggest themselves.

We could change our society profoundly by dropping the assumption that less than a tenth of the population is suitable to be taught basic concepts in maths and physics that have very wide application to our culture, such as normal distributions and conditional probability. This requires improving basic maths 5-16 and it also requires new courses in schools.

One of the things that we did in the DfE to do this was work with Fields Medallist Tim Gowers on a sort of ‘Maths for Presidents’ course. Professor Gowers wrote a fascinating blog on this course which you can read HERE. The DfE funded MEI to develop the blog into a real course. This has happened and the course is now being developed in schools. Physics for Future Presidents already exists and is often voted the most popular course at UC Berkeley (Cf. HERE). School-age pupils, arts graduates, MPs, and many Whitehall decision-makers would greatly benefit from these two courses.

We also need new inter-disciplinary courses in universities. For example, Oxford could atone for PPE by offering Ancient and Modern History, Physics for Future Presidents, and How to Run a Start Up. Such courses should connect to the work of Tetlock on The Good Judgement Project, as described above (I will return to this subject).

Other countries have innovated successfully in elite education. For example, after the shock of the Yom Kippur War, Israel established the ‘Talpiot’ programme which ‘aims to provide the IDF and the defense establishment with exceptional practitioners of research and development who have a combined understanding in the fields of security, the military, science, and technology. Its participants are taught to be mission-oriented problem-solvers. Each year, 50 qualified individuals are selected to participate in the program out of a pool of over 7,000 candidates. Criteria for acceptance include excellence in physical science and mathematics as well as an outstanding demonstration of leadership and character. The program’s training lasts three years, which count towards the soldiers’ three mandatory years of service. The educational period combines rigorous academic study in physics, computer science, and mathematics alongside intensive military training… During the breaks in the academic calendar, cadets undergo advanced military training… In addition to the three years of training, Talpiot cadets are required to serve an additional six years as a professional soldier. Throughout this period, they are placed in assorted elite technological units throughout the defense establishment and serve in central roles in the fields of research and development’ (IDF, 2012). The programme has also helped the Israeli hi-tech economy.****

If politicians had some basic training in mathematical reasoning, they could make better decisions amid complexity. If politicians had more exposure to the skills of a Bill Gates or Peter Thiel, they would be much better able to get things done.

I will explore the issue of training for politicians in a future blog.

Please leave corrections and comments below.

* It is very important to realise when the system one is examining is well approximated by a normal distribution and when by a power law. For example… When David Viniar (Goldman Sachs CFO) said of the 2008 financial crisis, ‘We were seeing things that were 25-standard-deviation events, several days in a row,’ he was discussing financial prices as if they can be accurately modelled by a normal distribution, and implying that events that should happen once every 10¹³⁵ years (the Universe is only ~1.4×10¹⁰ years old) were occurring ‘several days in a row’. He was either ignorant of basic statistics (unlikely) or taking advantage of the statistical ignorance of his audience. Actually, we have known for a long time that financial prices are not well modelled using normal distributions because they greatly underestimate the likelihood of bubbles and crashes. If politicians don’t know what ‘standard deviation’ means, it is obviously impossible for them to contribute much to detailed ideas on how to improve bank regulation. It is not hard to understand standard deviation and there is no excuse for this situation to continue for another generation.

** However, there is also a danger in the use of statistical models based on ‘big data’ analysis – ‘overfitting’ models and wrongly inferring a ‘signal’ from what is actually ‘noise’. We usually a) have a noisy data set and b) an inadequate theoretical understanding of the system, so we do not know how accurately the data represents some underlying structure (if there is such a structure). We have to infer a structure despite these two problems. It is easy in these circumstances to ‘overfit’ a model – to make it twist and turn to fit more of the data than we should, but then we are fitting it not to the signal but to the noise. ‘Overfit’ models can seem to explain more of the variance in the data – but they do this by fitting noise rather than signal (Silver, op. cit).

This error is seen repeatedly in forecasting, and can afflict even famous scientists. For example, Freeman Dyson tells a short tale about how, in 1953, he trekked to Chicago to show Fermi the results of a new physics model for the strong nuclear force. Fermi dismissed his idea immediately as having neither ‘a clear physical picture of the process that you are calculating’ nor ‘a precise and self-consistent mathematical formalism’. When Dyson pointed to the success of his model, Fermi quoted von Neumann, ‘With four parameters I can fit an elephant, and with five I can make him wiggle his trunk’, thus saving Dyson from wasting years on a wrong theory (A meeting with Enrico Fermi, by Freeman Dyson). Imagine how often people who think they have a useful model in areas not nearly as well-understood as nuclear physics lack a Fermi to examine it carefully.

There have been eleven recessions since 1945 but people track millions of statistics. Inevitably, people will ‘overfit’ many of these statistics to model historical recessions then ‘predict’ future ones. A famous example is the Superbowl factor. For 28 years out of 31, the winner of the Superbowl correctly ‘predicted’ whether the stock exchange rose or fell. A standard statistical test ‘would have implied that there was only about a 1-in-4,700,000 possibility that the relationship had emerged from chance alone.’ Just as someone will win the lottery, some arbitrary statistics will correlate with the thing you are trying to predict just by chance (Silver)

*** Many of these wrong forecasts were because the events were ‘out of sample’. What does this mean? Imagine you’ve taken thousands of car journeys and never had a crash. You want to make a prediction about your next journey. However, in the past you have never driven drunk. This time you are drunk. Your prediction is therefore out of sample. Predictions of US housing data were based on past data but there was no example of such huge leveraged price rises in the historical data. Forecasters who looked at Japan’s experience in the 1980’s better realised the danger. (Silver)

**** The old Technical Faculty of the KGB Higher School (rebaptised after 1991) ran similar courses; one of its alumni is Yevgeny Kaspersky, whose company first publicly warned of the cyberweapons Stuxnet and Flame (and who still works closely with his old colleagues). It would be interesting to collect information on elite intelligence and special forces training programmes. E.g. Post-9/11, US special forces (acknowledged and covert) have greatly altered including adding intelligence roles that were previously others’ responsibility or regarded as illegal for DOD employees. How does what is regarded as ‘core training’ for such teams vary, how is it changing, and why are some better than others at decisions under pressure and surviving disaster?

August 19, 2014 by dominiccummings

‘Standin’ by the window, where the light is strong’: de-extinction, machine intelligence, the search for extra-solar life, autonomous drone swarms bombing Parliament, genetics & IQ, science & politics, and much more @ SciFoo 2014

‘SciFoo’ 8-10 August 2014, the Googleplex, Silicon Valley, California.

On Friday 8 August, I woke up in Big Sur (the coast of Northern California), looked out over the waves breaking on the wild empty coastline, munched a delicious Mexican breakfast at Deetjen’s, then drove north on Highway 1 towards Palo Alto where a few hours later I found myself looking through the windows of Google’s HQ at a glittering sunset in Silicon Valley.

I was going to ‘SciFoo’. SciFoo is a weekend science conference. It is hosted by Larry Page at Google’s HQ in Silicon Valley and organised by various people including the brilliant Timo Hannay from Digital Science.

I was invited because of my essay that became public last year (cf. HERE). Of the 200+ people, I was probably the only one who made zero positive contribution to the fascinating weekend and therefore wasted a place, so although it was a fantastic experience for me the organisers should not invite me back and I feel guilty about the person who could not go because I was there. At least I can let others know about some of the things discussed… (Although it was theoretically ‘on the record unless stated otherwise’, I could tell that many scientists were not thinking about this and so I have left out some things that I think they would not want attributed. Given they were not experienced politicians being interviewed but scientists at a scientific conference, I’m erring on the side of caution, particularly given the subjects discussed.)

It was very interesting to see many of the people whose work I mentioned in my essay and watch them interacting with each other – intellectually and psychologically / physically.

I will describe some of the things that struck me though, because there are about 7-10 sessions going on simultaneously, this is only a small snapshot.

In my essay, I discuss some of the background to many of these subjects so I will put references [in square brackets] so people can refer to it if they want.

Please note that below I am reporting what I think others were saying – unless it is clear, I am not giving my own views. On technical issues, I do not have my ‘own’ views – I do not have relevant skills. All I can do is judge where consensus lies and how strong it is. Many important issues involve asking at least 1) is there a strong scientific consensus on X among physical scientists with hard quantitative data to support their ideas (uber-example, the Standard Model of particle physics), b) what are the non-science issues, such as ‘what will it cost, who pays/suffers and why?’ On A, I can only try to judge what technically skilled people think. B is a different matter.

Whether you were there or not, please leave corrections / additions / questions in the comments box. Apologies for errors…

In a nutshell, a few likely scenarios / ideas, without spelling out caveats… 1) Extinct species are soon going to be brought back to life and the same technology will be used to modify existing species to help prevent them going extinct. 2) CRISPR – a new gene editing technology – will be used to cure diseases and ‘enhance’ human performance but may also enable garage bio-hackers to make other species extinct. 3) With the launch of satellites in 2017/18, we may find signs of life by 2020 among the ~10¹¹ exoplanets we now know exist just in our own galaxy though it will probably take 20-30 years, but the search will also soon get crowdsourced in a way schools can join in. 4) There is a reasonable chance we will have found many of the genes for IQ within a decade via BGI’s project, and the rich may use this information for embryo selection. 5) ‘Artificial neural networks’ are already outperforming humans on various pattern-recognition problems and will continue to advance rapidly. 6) Automation will push issues like a negative income tax onto the political agenda as millions lose their jobs to automation. 7) Autonomous drones will be used for assassinations in Europe and America shortly. 8) Read Neil Gershenfeld’s book ‘FAB’ if you haven’t and are interested in science education / 3D printing / computer science (or at least watch his TED talks). 9) Scientists are desperate to influence policy and politics but do not know how.

Biological engineering / computational biology / synthetic biology [Section 4]

George Church (Harvard), a world-leading biologist, spoke at a few sessions and his team’s research interests were much discussed. (Don’t assume he said any specific thing below.)

The falling cost of DNA sequencing continues to spur all sorts of advances. It has fallen from a billion dollars per genome a decade ago to less than a thousand dollars now (a million-fold improvement), and the Pentagon is planning on it reaching $100 soon. We can also sequence cancer cells to track their evolution in the body.

CRISPR. CRISPR is a new (2012) and very hot technology that is a sort of ‘cut and paste’ gene editing tool. It allows much more precise and effective engineering of genomes. Labs across America are rushing to apply it to all sorts of problems. In March this year, it was used to correct faulty genes in mice and cure them of a liver condition. It plays a major part in many of the biological issues sketched below.

‘De-extinction’ (bringing extinct species back to life). People are now planning the practical steps for de-extinction to the extent that they are scoping out land in Siberia where woolly mammoths will roam. As well as creating whole organisms, they will also grow organs modified by particular genes to test what specific genes and combinations do. This is no longer sci-fi – it is being planned and is likely to happen. The buffalo population was recently re-built (Google serves buffalo burgers in its amazing kitchens) from a tiny population to hundreds of thousands and there seems no reason to think it is impossible to build a significant population from scratch.

What does this mean? You take the DNA from an animal, say a woolly mammoth buried in the ground, sequence it, then use the digitised genome to create an embryo and either grow it in a similar animal (e.g. elephant for a mammoth) or in an artificial womb. (I missed the bit explaining the rationale for some of the proposed projects but, apart from the scientific reasons, one rationale for the mammoth was described as a conservation effort to preserve the frozen tundra and prevent massive amounts of greenhouse gases being released from beneath it.)

There are also possibilities of using this technology for conservation. For example, one could re-engineer the Asian elephant so that it could survive in less hospitable climates (e.g. modify the genes that produce haemoglobin so it is viable in colder places).

Now that we have sequenced the genome for Neanderthals (and learned that humans interbred with them, so you have traces of their DNA – unless you’re an indigenous sub-Saharan African), there is no known physical reason why we could not bring a Neanderthal back to life once the technology has been refined on other animals. This obviously raises many ethical issues – e.g. if we did it, they would have to be given the same legal rights as us (one distinguished person said that if there were one in the room with us we would not notice, contra the pictures often used to illustrate them). It is assumed by many that this will happen (nobody questioned the assumption) – just as it seemed to be generally assumed that human cloning will happen – though probably not in a western country but somewhere with fewer legal restrictions, after the basic technologies have been refined. (The Harvard team gets emails from women volunteering to be the Neanderthal’s surrogate mum.)

‘Biohacking’. Biohacking is advancing faster than Moore’s Law. CRISPR editing will allow us to enhance ourselves. E.g. Tibetans have evolved much more efficient systems for coping with high altitude, and some Africans have much stronger bones than the rest of us (see below). Will we reengineer ourselves to obtain these advantages? CRISPR obviously also empowers all sorts of malevolent actors too – cf. this very recent paper (by Church et al). It may soon be possible for people in their garages to edit genomes and accidentally or deliberately drive species to extinction as well as attempt to release deadly pathogens. I could not understand why people were not more worried about this – I hope I was missing a lot. (Some had the attitude that ‘nature already does bio-terrorism’ so we should relax. I did not find this comforting and I’m sure I am in the majority so for anybody influential reading this I would strongly advise you not to use this argument in public advocacy or it is likely to accelerate calls for your labs to be shut down.)

‘Junk’. There is more and more analysis of what used to be called ‘junk DNA’. It is now clear that far from being ‘junk’ much of this has functions we do not understand. This connects to the issue that although we sequenced the human genome over a decade ago, the quality of the ‘reference’ version is not great and (it sounded like from the discussions) it needs upgrading.

‘Push button’ cheap DNA sequencers are around the corner. Might such devices become as ubiquitous as desktop printers? Why doesn’t someone create a ‘gene web browser’ that can cope with all the different data formats for genomes?

Privacy. There was a lot of talk about ‘do you want your genome on the web?’. I asked a quick informal pop quiz (someone else’s idea): there was unanimity that ‘I’d much rather my genome was on the web than my browsing history’. [UPDATE: n<10 and perhaps they were tongue in cheek!? One scientist pointed out in a session that when he informed his insurance company, after sequencing his own genome, that he had a very high risk of getting colon cancer, they raised his premiums. There are all sorts of reasons one would want to control genomic information and I was being a bit facetious.]

In many ways, computational biology and synthetic biology have that revolutionary feeling of the PC revolution in the 1970s – huge energy, massive potential for people without big resources to make big contributions, the young crowding in, the feeling of dramatic improvements imminent. Will this all seem ‘too risky’? It’s hard to know how the public will respond to risk. We put up with predictable annual carnage from car accidents but freak out over trivia. We ignore millions of deaths in the Congo but freak out over a handful in Israel/Gaza. My feeling is some of the scientists are too blasé about how the public will react to the risks, but I was wrong about how much fear there would be about the news that scientists recently deliberately engineered a much more dangerous version of an animal flu.

AI / machine learning / neuroscience [Section 5].

Artificial neural networks (NNs), now often referred to as ‘deep learning’, were first created 50 years ago but languished for a while when progress slowed. The field is now hot again. (Last year Google bought some companies leading the field, and a company, Boston Dynamics, that has had a long-term collaboration with DARPA.)

Jurgen Schmidhuber explained progress and how NNs have recently approached or surpassed human performance in various fields. E.g. recently NNs have surpassed human performance in recognising traffic signals (0.56% error rate for the best NN versus 1.16% for humans). Progress in all sorts of pattern recognition problems is clearly going to continue rapidly. E.g. NNs are now being used to automate a) the analysis of scans for cancer cells and b) the labelling of scans of human brains – so artificial neural networks are now scanning and labelling natural neural networks.

Steve Hsu has blogged about this session here:

http://infoproc.blogspot.co.uk/2014/08/neural-networks-and-deep-learning.html?m=1

Michael Nielsen is publishing an education project online for people to teach themselves the basics of neural networks. It is brilliant and I would strongly advise teachers reading this blog to consider introducing it into their schools and doing the course with the pupils.

http://neuralnetworksanddeeplearning.com

Neil Gershenfeld (MIT) gave a couple of presentations. One was on developments in computer science connecting: non-‘von Neumann architecture’, programmable matter, 3D printing, ‘the internet of things’ etc. [Cf. Section 3.] NB. IBM announced this month substantial progress in their quest for a new computer architecture that is ‘non-Von Neumann’: cf. this –

http://venturebeat.com/2014/08/07/ibms-synapse-marshals-the-power-of-the-human-brain-in-a-computer/view-all/

Another was on the idea of an ‘interspecies internet’. We now know many species can recognise each other, think, and communicate much better than we realised. He showed bonobos playing music with Peter Gabriel and dolphins communicating. He and others are plugging them into the internet. Some are doing this to help the general goal of figuring out how we might communicate with intelligent aliens – or how they might communicate with us.

(Gershenfeld’s book FAB led me to push 3D printing into the new National Curriculum and I would urge school science teachers to watch his TED talks and read this book. [INSERTED LATER: Some people have asked about this point. I (I thought obviously) did not mean I wrote the NC document. I meant – I pushed the subject into the discussions with the committees/drafters who wrote the NC. Experts in the field agreed it belonged. When it came out, this was not controversial. We also funded pilots with 3D printers so schools could get good advice about how to teach the subject well.] His point about 3D printers restoring the connection between thinking and making – lost post-Renaissance – is of great importance and could help end the foolishly entrenched ‘knowledge’ vs ‘skills’ and academic vs vocational trench wars. Gove actually gave a speech about this not long before he was moved and as far as I could tell it got less coverage than any speech he ever gave, thus proving the cliché about speeches on ‘skills’.)

There were a few presentations about ‘computational neuroscience’. I could not understand anything much as they were too technical. It was clear that there is deep concern among EU neuroscientists about the EU’s huge funding for Henry Markram’s Human Brain Project. One leading neuroscientist said to me that the whole project is misguided as it does not have clear focused goals and the ‘overhype’ will lead to public anger in a few years. Apparently, the EU is reconsidering the project and its goals. I have no idea about the merits of these arguments. I have a general prejudice that, outside special circumstances, experience suggests that it is better to put funding into many pots and see what works, as DARPA does.

There are all sorts of crossovers between: AI / neuroscience / big data / NNs / algorithmic pattern recognition in other fields.

Peter Norvig, a leader in machine intelligence, said that he is more worried about the imminent social implications of continued advances making millions unemployed than he is about a sudden ‘Terminator / SKYNET’ scenario of a general purpose AI bootstrapping itself to greater than human intelligence and exterminating us all. Let’s hope so. It is obvious that this field is going to keep pushing boundaries – in open, commercial, and classified projects – so we are essentially going to be hoping for the best as we make more and more advances in AI. The idea of a ‘negative income tax’ – or some other form of essentially paying people X just to live – seems bound to return to the agenda. I think it could be a way around all sorts of welfare arguments. The main obstacle, it seems to me, is that people won’t accept paying for it if they think uncontrolled immigration will continue as it is now.

Space [Section 2]

There was great interest in various space projects and some senior people from NASA. There is much sadness at how NASA, despite many great people, has become a normal government institution – ie. caught in DC politics, very bureaucratic, and dysfunctional in various ways. On the other hand, many private ventures are now growing. E.g. Elon Musk is lowering the $/kg of getting material into orbit and planning a non-government Mars mission. As I said in my essay, really opening up space requires a space economy – not just pure science and research (such as putting telescopes on the far side of the moon, which we obviously should do). Columbus opened up America – not the Vikings.

There is another obvious motive. As Carl Sagan said, if the dinosaurs had had a space programme, they’d still be here. In the long-term we either develop tools for dealing with asteroids or we will be destroyed. We know this for sure. I think I heard that NASA is planning to park a small asteroid close to the moon around 2020 but I may have misheard / misunderstood.

Mario Livio led a great session on the search for life on exoplanets. The galaxy has ~10¹¹ stars and there is ~1 planet on average per star. There are ~10¹¹ galaxies, so a Fermi estimate is there are ~10²² planets – 10 billion trillion planets – in the observable universe (this number is roughly 1,000 times bigger than the number you get in the fable of putting a grain of rice on the first square of a chessboard and doubling on each subsequent square). Many of them are in the ‘habitable zone’ around stars.

In 2017/18, there are two satellites launching that will be able to do spectroscopy on exoplanets – i.e. examine their atmospheres and detect things like oxygen and water. ‘If we get lucky’, these satellites will find ‘bio-signatures’ of life. If they find life having looked at only a few planets, then it would mean that life is very common. ‘More likely’ is it will take 20-30 years and a new generation of space-based telescopes to find life. If planets are found with likely biosignatures, then it would make sense to turn SETI’s instruments towards them to see if they find anything. (However, we are already phasing out the use of radio waves for various communications – perhaps the use of radio waves is only a short window in the lifetime of a civilisation.) There are complex Bayesian arguments about what we might infer about our own likely future given various discoveries but I won’t go into those now. (E.g. if we find life is common but no traces of intelligent life, does this mean a) the evolution of complex life is not a common development from simple life; b) intelligent life is also common but it destroys itself; c) they’re hiding, etc.)

A very impressive (and helpful towards the ignorant like me) young scientist working on exoplanets called Oliver Guyon demonstrated a fascinating project to crowdsource the search for exoplanets by building a global network of automated cameras – PANOPTES (www.projectpanoptes.org). His team has built a simple system that can find exoplanets using normal digital cameras costing less than $1,000. They sit in a box connected to a 12V power supply, automatically take pictures of the night sky every few seconds, then email the data to the cloud. There, the data is aggregated and algorithms search for exoplanets. These units are cheap (can’t remember what he said but I think <$5,000). Everything is open-source, open-hardware. They will start shipping later this year and will make a brilliant school science project. Guyon has made the project with schools in mind so that assembling and operating the units will not require professional level skills. They are also exploring the next move to connect smartphone cameras.

Building the >15m diameter space telescopes we need to search for life seems to me an obvious priority for scientific budgets – it is one of the handful of the most profound questions facing us.

There was an interesting cross-over discussion about ‘space and genetics’ in which people discussed various ways in which space exploration would encourage / require genetic modification. E.g.1 some sort of rocket fuel has recently been discovered to exist in large quantities on Mars. This is very handy but the substance is toxic. It might therefore make sense to modify humans going to live on Mars to be resistant. E.g.2 Space travel weakens bones. It has been discovered that mutations in the human population can improve bone strength by 8 standard deviations. This is a massive improvement – for comparison, 8 SDs in IQ covers people from severely mentally disabled to Nobel-winners. This was discovered by a team of scientists in Africa who noticed that people in a local tribe who got hit by cars did not suffer broken bones, so they sequenced the locals’ genomes. (Someone said there have already been successful clinical trials testing this discovery in a real drug to deal with osteoporosis.) E.g.3 Engineering E. Coli shows that just four mutations can improve resistance to radiation by ?1,000 times (can’t read my note).

Craig Venter and others are thinking about long-term projects to send ‘von Neumman-bots’ (self-replicating space drones) across the universe containing machines that could create biological life once they arrive somewhere interesting, thus avoiding the difficult problems of keeping humans alive for thousands of years on spaceships. (Nobel-winning physicist Gerard t’ Hooft explains the basic principles of this in his book Playing with planets.)

This paper (August 2014) summarises issues in the search for life:

http://www.pnas.org/content/early/2014/08/01/1304213111.full.pdf

Finding the genes for IQ and engineering possibilities [Section 5].

When my essay came out last year, there was a lot of mistaken reporting that encouraged many in the education world to grab the wrong end of the stick about IQ, though the BBC documentary about the controversy (cf. below) was excellent and a big step forward. It remains the case that very few people realise that in the last couple of years direct examination of DNA has now vindicated the consistent numbers on IQ heritability from decades of twin/adoption studies.

The rough heritability numbers for IQ are no longer in doubt among physical scientists who study this field: it is roughly 50% heritable at age ~18-20 and this number rises towards 70-80% for older adults. This is important because IQ is such a good predictor of the future – it is a better predictor than social class. E.g. The long-term Study of Mathematically Precocious Youth, which follows what has happened to children with 1:10,000 ability, shows among many things that a) a simple ‘noisy’ test administered at age 12-13 can make amazingly accurate predictions about their future, and b) achievements such as scientific breakthroughs correlate strongly with IQ. (If people looked at the data from SMPY, then I think some of the heat and noise in the debate would fade but it is a sad fact that approximately zero senior powerful people in the English education world had even heard of this study before the furore over Plomin last year.)

Further, the environmental effects that are important are not the things that people assume. If you test the IQ of an adopted child in adulthood and the parents who adopted it, you find approximately zero correlation – all those anguished parenting discussions had approximately no measurable impact on IQ. (This does not mean that ‘parenting doesn’t matter’ – parents can transfer narrow skills such as playing the violin.) In the technical language, the environmental effects that are important are ‘non-shared’ environmental effects – i.e. they are things that two identical twins do not experience in the same way. We do not know what they are. It is reasonable to think that they are effectively random tiny events with nonlinear effects that we may never be able to track in detail – cf. this paper for a discussion of this issue in the context of epidemiology: http://ije.oxfordjournals.org/content/40/3/537.full.pdf+html

There remains widespread confusion on this subject among social scientists, education researchers, and the worlds of politics and the media where people were told misleading things in the 1980s and 1990s and do not realise that the debates have been transformed. To be fair, however, it was clear from this weekend that even many biologists do not know about new developments in this field so it is not surprising that political journalists and education researchers do not.

(An example of confusion in the political/media world… In my essay, I used the technical term ‘heritable’ which is a population statistic – not a statement about an individual. I also predicted that media coverage would confuse the subject (e.g. by saying things like ‘70% of your IQ comes from genes’). Sure enough some journalists claimed I said the opposite of what I actually said then they quoted scientists attacking me for making a mistake that not only did I not make but which I actually warned about. Possibly the most confused sentence of all those in the media about my essay was the line ‘wealth is more heritable than genes’, which was in Polly Toynbee’s column and accompanying headline in the Guardian. This sentence is a nonsense sentence as it completely mangles the meaning of the term ‘heritable’. Much prominent commentary from politicians and sociologists/economists on ‘social mobility’ is gibberish because of mistaken assumptions about genes and environment. The Endnote in my essay has links to work by Plomin, Hsu et al that explains it all properly. This interview with Plomin is excellent: http://www.spectator.co.uk/features/8970941/sorry-but-intelligence-really-is-in-the-genes/. This recent BBC radio programme is excellent and summarises the complex issues well: http://www.bbc.co.uk/programmes/b042q944/episodes/guide)

I had a fascinating discussion/tutorial at SciFoo with Steve Hsu. Steve Hsu is a professor of theoretical physics (and successful entrepreneur) with a long interest in IQ (he also runs a brilliant blog that will keep you up to speed on all sorts). He now works part time on the BGI project in China to discover the genes responsible for IQ.

IQ is very similar to height from the perspective of behavioural genetics. Height has the advantage that it is obviously easier to measure than IQ but it has roughly the same heritability. Large scale GWAS are already identifying some of the genes responsible for height. Hsu recently watched a talk by Fields Medallist Terry Tao and realised that a branch of maths could be used to examine the question – how many genomes do we need to scan to identify a substantial number of the genes for IQ? His answer: ‘roughly 10k moderately rare causal variants of mostly negative effect are responsible for normal population variation’ and finding them will require sequencing roughly a million genomes. The falling cost of sequencing DNA means that this is within reach. ‘At the time of this writing SNP genotyping costs are below $50 USD per individual, meaning that a single super-wealthy benefactor could independently fund a crash program for less than $100 million’ (Hsu).

The BGI project to find these genes has hit some snags recently (e.g. a US lawsuit between the two biggest suppliers of gene sequencing machines). However, it is now expected to start again soon. Hsu thinks that within a decade we could find many of the genes responsible for IQ. He has just put his fascinating paper on this subject on his blog (there is also a Q&A on p.27 that will be very useful for journalists):

http://infoproc.blogspot.co.uk/2014/08/genetic-architecture-of-intelligence.html

Just discovering a substantial fraction of the genes would be momentous in itself but there is more. It is already the case that farmers use genomes to make predictions about cows’ properties and behaviour (‘genotype to phenotype’ predictions). It is already the case that rich people could use in vitro fertilisation to select the egg which they think will be most advantageous, because they can sequence genomes of multiple eggs and examine each one to look for problems then pick the one they prefer. Once we identify a substantial number of IQ genes, there is no obvious reason why rich people will not select the egg that has the highest prediction for IQ.

This clearly raises many big questions. If the poor cannot do the same, then the rich could quickly embed advantages and society could become not only more unequal but also based on biological classes. One response is that if this sort of thing does become possible, then a national health system should fund everybody to do this. (I.e. It would not mandate such a process but it would give everybody a choice of whether to make use of it.) Once the knowledge exists, it is hard to see what will stop some people making use of it and offering services to – at least – the super-rich.

It is vital to separate two things: a) the basic science of genetics and cognition (which must be allowed to develop), and b) the potential technological applications and their social implications. The latter will rightly make people deeply worried, given our history, and clearly require extremely serious public debate. One of the reasons I wrote my essay was to try to stimulate such debate on the biggest – and potentially most dangerous – scientific issues. By largely ignoring such issues, Westminster, Whitehall, and the political media are wasting the time we have to discuss them so technological breakthroughs will be unnecessarily shocking when they come.

Hsu’s contribution to this research – and his insight when listening to Tao about how to apply a branch of mathematics to a problem – is also a good example of how the more abstract fields of maths and physics often make contributions to the messier study of biology and society. The famous mathematician von Neumann practically invented some new fields outside maths and made many contributions to others. The physicist-mathematician Freeman Dyson recently made a major contribution to Game Theory which had lain unnoticed for decades until he realised that a piece of maths could be applied to uncover new strategies (Google “Dyson zero determinant strategies” and cf. this good piece: http://www.americanscientist.org/issues/id.16112,y.0,no.,content.true,page.1,css.print/issue.aspx).

However, this also raises a difficult issue. There is a great deal of Hsu’s paper – and the subject of IQ and heritability generally – that I do not have the mathematical skills to understand. This will be true of a large fraction of education researchers in education departments – I would bet a large majority. This problem is similar for many other vital issues (and applies to MPs and their advisers) and requires general work on translating such research into forms that can be explained by the media.

Kathryn Ashbury also did a session on genes and education but I went to a conflicting one with George Church so unfortunately I missed it.

‘Big data’, simulations, and distributed systems [Section 6&7]

The rival to Markram’s Brain Project for mega EU funding was Dirk Helbing (ETH Zurich) and his project for new simulations to aid policy-making. Helbing was also at SciFoo and gave a couple of presentations. I will write separately about this.

Helbing says convincingly: ‘science must become a fifth pillar of democracies, besides legislation, executive, jurisdiction, and the public media’. Many in politics hope that technology will help them control things that now feel out of control. This is unlikely. The amount of data is growing at a faster rate than the power of processing and the complexity of networked systems grows factorially therefore top-down control will become less and less effective.

The alternative? ‘Distributed (self-)control, i.e. bottom-up self-regulation’. E.g. Helbing’s team has invented self-regulating traffic lights driven by traffic flows that can ‘outperform the classical top-down control by a conventional traffic center.’

‘Can we transfer and extend this principle to socio-economic systems? Indeed, we are now developing mechanisms to overcome coordination and cooperation failures, conflicts, and other age-old problems. This can be done with suitably designed social media and sensor networks for real-time measurements, which will eventually weave a Planetary Nervous System. Hence, we can finally realize the dream of self-regulating systems… [S]uitable institutions such as certain social media – combined with suitable reputation systems – can promote other-regarding decision-making. The quick spreading of social media and reputation systems, in fact, indicates the emergence of a superior organizational principle, which creates collective intelligence by harvesting the value of diversity…’

His project’s website is here:

http://www.futurict.eu

I wish MPs and spads in all parties would look at this project and Helbing’s work. It provides technologically viable and theoretically justifiable mechanisms to avoid the current sterile party debates about delivery of services. We must move from Whitehall control to distributed systems…

Science and politics

Unsurprisingly, there was a lot of grumbling about politicians, regulation, Washington gridlock, bureaucracy and so on.

Much of it is clearly justified. Some working in genetics had stories about how the regulations forbid them to tell people about imminently life threatening medical problems they discover. Others were bemoaning the lack of action on asteroid defence and climate change.

Some of these problems are inherently extremely difficult, as I discuss in my essay. On top of this, though, is the problem that many (most?) scientists do not know how to go about changing things.

It was interesting that some very eminent scientists, all much cleverer than ~100% of those in politics [INSERT: better to say ‘all with higher IQ than ~100% of those in politics’], have naive views about how politics works. In group discussions, there was little focused discussion about how they could influence politics better even though it is clearly a subject that they care about very much. (Gershenfeld said that scientists have recently launched a bid to take over various local government functions in Barcelona, which sounds interesting.)

A few times I nearly joined in the discussion but I thought it would disrupt things and distract them. In retrospect, I think this may have been a mistake and I should have spoken up. But also I am not articulate and I worried I would not be able to explain their errors and it would waste their time.

I will blog on this issue separately. A few simple observations…

To get things changed in politics, scientists need mechanisms a) to agree priorities in order to focus their actions on b) roadmaps with specifics. Generalised whining never works. The way to influence politicians is to make it easy for them to fall down certain paths without much thought, and this means having a general set of goals but also a detailed roadmap the politicians can apply, otherwise they will drift by default to the daily fog of chaos and moonlight.

Scientists also need to be prepared to put their heads above the parapet and face controversy. Many comments amounted to ‘why don’t politicians do the obviously rational thing without me having to take a risk of being embroiled in media horrors’. Sorry guys but this is not how it works.

Many academics are entirely focused on their research and do not want to lose time to politics. This is entirely reasonable. But if you won’t get involved you can have little influence other than lending your name to the efforts of others.

Working in the Department for Education, I have experienced in England that very few scientists were prepared to face controversy over the issue of A Levels (exams at 18) and university entry / undergraduate standards even though this problem directly affected their own research area. Many dozens sought me out 2007-14 to complain about existing systems. I can count on the fingers of one hand those who rolled the dice and did things in the public domain that could have caused them problems. I have heard many scientists complain about media reports but when I’ve said – ‘write a blog explaining why they’re wrong’, the answer is almost invariably ‘oh, the VC’s office would go mad’. If they won’t put their heads above the parapet on an issue that directly touches their own subject and career, how much are they likely to achieve in moving political debate in areas outside their own fields?

Provided scientists a) want to avoid controversy and b) are isolated, they cannot have the leverage they want. The way to minimise controversy is to combine in groups – for the evolutionary biologists reading this, think SHOALS! – so that each individual is less exposed. But you will only join a shoal if you agree a common purpose.

I’m going to do a blog on ‘How scientists can learn from Bismarck and Jean Monnet to influence politics‘. Monnet avoided immediate battles for power in favour of ‘preparing the future’ – i.e. having plans in his pocket for when crises hit and politicians were desperate. He created the EEC in this way. In the same way people find it extremely hard to operationalise the lessons of Thucydides or Bismarck, they do not operationalise the lessons from Monnet. It would be interesting if scientists did this in a disciplined way. In some ways, it seems to me vital if we are to avoid various disasters. It is also necessary, however, to expose scientists to the non-scientific factors in play.

Anyway, it would be worth exploring this question: can very high IQ people with certain personality traits (like von Neumann, not like Gödel) learn enough in half a day’s exposure to case studies of successful political action to enable them to change something significant in politics, provided someone else can do most of the admin donkey work? I’m willing to bet the answer is YES. Whether they will then take personal risks by ACTING is another question.

A physicist remarked: ‘we’re bitching about politicians but we can’t even sort out our own field of scientific publishing which is a mess’.

NB. for scientists who haven’t read anything I’ve read before, do not make the mistake of thinking I am defending politicians. If you read other stuff I’ve written you will see that I have made all the criticisms that you have. But that doesn’t mean that scientists cannot do much better than they are at influencing policy.

A few general comments

1. It has puzzled me for over a decade that a) one of the few things the UK still has that is world class is Oxbridge, b) we have the example of Silicon Valley and our own history of post-1945 bungling to compare it with (e.g. how the Pentagon treated von Neumann and how we treated Turing viz the issue of developing computer science), yet c) we persistently fail to develop venture capital-based hubs around Oxbridge on the scale they deserve. As I pottered down University Avenue in Palo Alto looking for a haircut, past venture capital offices that can provide billions in start-up investment, I thought: you’ve made a few half-hearted attempts to persuade people to do more on this, when you get home try again. So I will…

2. It was interesting to see how physicists have core mathematical skills that allow them to grasp fundamentals of other fields without prior study. Watching them reminded me of Mandelbrot’s comment that:

‘It is an extraordinary feature of science that the most diverse, seemingly unrelated, phenomena can be described with the same mathematical tools. The same quadratic equation with which the ancients drew right angles to build their temples can be used today by a banker to calculate the yield to maturity of a new, two-year bond. The same techniques of calculus developed by Newton and Leibniz two centuries ago to study the orbits of Mars and Mercury can be used today by a civil engineer to calculate the maximum stress on a new bridge… But the variety of natural phenomena is boundless while, despite all appearances to the contrary, the number of really distinct mathematical concepts and tools at our disposal is surprisingly small… When we explore the vast realm of natural and human behavior, we find the most useful tools of measurement and calculation are based on surprisingly few basic ideas.’

3. High status people have more confidence in asking basic / fundamental / possibly stupid questions. One can see people thinking ‘I thought that but didn’t say it in case people thought it was stupid and now the famous guy’s said it and everyone thinks he’s profound’. The famous guys don’t worry about looking stupid and they want to get down to fundamentals in fields outside their own.

4. I do not mean this critically but watching some of the participants I was reminded of Freeman Dyson’s comment:

‘I feel it myself, the glitter of nuclear weapons. It is irresistible if you come to them as a scientist. To feel it’s there in your hands. To release the energy that fuels the stars. To let it do your bidding. And to perform these miracles, to lift a million tons of rock into the sky, it is something that gives people an illusion of illimitable power, and it is in some ways responsible for all our troubles, I would say, this is what you might call ‘technical arrogance’ that overcomes people when they see what they can do with their minds.’

People talk about rationales for all sorts of things but looking in their eyes the fundamental driver seems to be – am I right, can I do it, do the patterns in my mind reflect something real? People like this are going to do new things if they can and they are cleverer than the regulators. As a community I think it is fair to say that outside odd fields like nuclear weapons research (which is odd because it still requires not only a large collection of highly skilled people but also a lot of money and all sorts of elements that are hard (but not impossible) for a non-state actor to acquire and use without detection), they believe that pushing the barriers of knowledge is right and inevitable. Fifteen years on from the publication by Silicon Valley legend Bill Joy of his famous essay (‘Why the future doesn’t need us’), it is clear that many of the things he feared have proceeded and there remains no coherent government approach or serious international discussion. (I am not suggesting that banning things is generally the way forward.)

5. The only field where there was a group of people openly lobbying for something to be made illegal was the field of autonomous lethal drones. (There is a remorseless logic that means that countermeasures against non-autonomous drones (e.g. GPS-spoofing) incentivises one to make one’s drones autonomous. They can move about waiting to spot someone’s face then destroy them without any need for human input.) However, the discussion confirmed my view that even if this might be a good idea – it is doomed, in the short-term at least. I wonder what is to stop someone sending a drone swarm across the river and bombing Parliament during PMQs. Given it will be possible to deploy autonomous drones anonymously, it seems there may be a new era of assassinations coming, apart from all the other implications of drones. Given one may need a drone swarm to defend against drone swarm, I can’t see them being outlawed any time soon. (Cf. Suarez’s Kill Decision for a great techno-thriller on the subject.)

(Also, I thought that this was an area where those involved in cutting edge issues could benefit from talking to historians. E.g. my understanding is that we filmed the use of anthrax on a Scottish island and delivered the footage to the Nazis with the message that we would anthrax Germany if they used chemical weapons – i.e. the lack of chemical warfare in WWII was a case of successful deterrence, not international law.)

6. A common comment is – ‘technology X [e.g. in vitro fertilisation] was denounced at the time but humans adapt to such changes amazingly fast, so technology Y will be just the same’. This is a reasonable argument in some ways but I cannot help but think that many will think de-extinction, engineered bio-weapons, or human clones are going to be perceived as qualitative changes far beyond things like in vitro fertilisation.

7. Daniel Suarez told me what his next techno-thriller is about but if I put it on my blog he will deploy an autonomous drone with face recognition AI to kill me, so I’m keeping quiet. If you haven’t read Daemon, read it – it’s a rare book that makes you laugh out loud about how clever the plot is.

8. Von Neumann was heavily involved not only in the Manhattan Project but also the birth of the modern computer, the creation of the hydrogen bomb, and nuclear strategy. Before his tragic early death, he wrote a brilliant essay about the political problem of dealing with advanced technology which should be compulsory reading for all politicians aspiring to lead. It summarises the main problems that we face – ‘for progress, there is no cure…’

http://features.blogs.fortune.cnn.com/2013/01/13/can-we-survive-technology/

As I said at the top, any participants please tell me where I went wrong, and thanks for such a wonderful weekend.

Complexity, ‘fog and moonlight’, prediction, and politics I

‘What can be avoided

Whose end is purposed by the mighty gods?

Yet Caesar shall go forth, for these predictions

Are to the world in general as to Caesar.’

Julius Caesar, II.2.

‘Ideas thus made up of several simple ones put together, I call Complex; such as are Beauty, Gratitude, a Man, an Army, the Universe.’ Locke.

‘I can calculate the motion of heavenly bodies but not the madness of people.’ Newton, after the South Sea Bubble ‘Ponzi scheme’.

‘It is a wonderful feeling to recognise the unity of complex phenomena that to direct observation appear to be quite separate things.’ Einstein to Grossman, 1901.

‘All stable processes we shall predict. All unstable processes we shall control.’ Von Neumann.

‘Imagine how much harder physics would be if electrons had feelings.’ Richard Feynman.

At the beginning of From Russia With Love (the movie not the book), Kronsteen, a Russian chess master and SPECTRE strategist, is summoned to Blofeld’s lair to discuss the plot to steal the super-secret ‘Lektor Decoder’ and kill Bond. Kronsteen outlines to Blofeld his plan to trick Bond into stealing the machine for SPECTRE.

Blofeld: Kronsteen, you are sure this plan is foolproof?

Kronsteen: Yes it is, because I have anticipated every possible variation of counter-move.

Political analysis is full of chess metaphors, reflecting an old tradition of seeing games as models of physical and social reality. (‘Time is a child moving counters in a game; the royal power is a child’s’, Heraclitus.) A game which has ten different possible moves at each turn and runs for two turns has 10² possible ways of being played; if it runs for fifty turns it has 10⁵⁰ possible ways of being played, ‘a number which substantially exceeds the number of atoms in the whole of our planet earth’ (Holland); if it runs for ninety turns it has 10⁹⁰ possible ways of being played, which is about the estimated number of atoms in the Universe. Chess is merely 32 pieces on an 8×8 grid with a few simple rules but the number of possible games is much greater than 10⁹⁰.

Many practical problems (e.g logistics, designing new drugs) are equivalent to the Travelling Salesman Problem (TSP). For any TSP involving travelling to n cities, the number of possible tours when starting with a specific city is: (n-1)!/2. For 33 cities, the total number of possible journeys is:

32!/2 = 131,565,418,466,846,765,083,609,006,080,000,000

The IBM Roadrunner, the fastest supercomputer in the world in 2009, could perform 1,457 trillion operations per second. If we could arrange the tours such that examining each one would take only one arithmetical operation, then it would take it ~28 trillion years to examine all possible routes between 33 cities, about twice the estimated age of the Universe. As n grows linearly (add one city, add another etc), the number of possible routes grows exponentially. The way in which the number of possible options scales up exponentially as the number of agents scales up linearly, and the difficulty of finding solutions quickly in vast search landscapes, connects to one of the most important questions in maths and computer science, the famous $1 million dollar ‘P=NP?’ Clay Millennium Prize.

Kronsteen’s confidence, often seen in politics, is therefore misplaced even in chess. It is far beyond our ability to anticipate ‘every possible variation of counter-move’ yet chess is simple compared to the systems that scientists or politicians have to try to understand and predict in order to try to control. These themes of uncertainty, nonlinearity, complexity and prediction have been ubiquitous motifs of art, philosophy, and politics. We see them in Homer, where the gift of an apple causes the Trojan War; in Athenian tragedy, where a chance meeting at a crossroads settles the fate of Oedipus; in Othello’s dropped handkerchief; and in War and Peace with Nikolai Rostov, playing cards with Dolohov, praying that one little card will turn out differently, save him from ruin, and allow him to go happily home to Natasha.

‘I know that men are persuaded to go to war in one frame of mind and act when the time comes in another, and that their resolutions change with the changes of fortune… The movement of events is often as wayward and incomprehensible as the course of human thought; and this is why we ascribe to chance whatever belies our calculation.’ Pericles to the Athenians.

Maths and models

Because of the ‘unreasonable effectiveness of mathematics’ in providing the ‘language of nature’ and foundations for a scientific civilization, we understand some systems very well and can make very precise predictions based on accurate quantitative models. Sometimes a mathematical model predicts phenomena which are later found (e.g. General Relativity’s field equations); sometimes an experiment reveals a phenomenon that awaits an effective mathematical model (e.g. the delay between the discovery of superconductivity and a quantum theory). The work of mathematicians on ‘pure’ problems has often yielded ideas that have waited to be rediscovered by physicists. The work of Euclid, Apollonius and Archimedes on ellipses would be used centuries later by Kepler for his theory of planetary motion. The work of Riemann on non-Euclidean four-dimensional geometry was (thanks to Grossmann) used by Einstein for General Relativity. The work of various people since the 16^th Century on complex numbers would be used by Heisenberg et al for quantum mechanics in the 1920s.

The work of Cantor, Gödel, and Turing (c. 1860-1936) on the logical foundations of mathematics, perhaps the most abstract and esoteric subject, gave birth to computers. The work of Galois on ‘groups’ (motivated by problems with polynomial equations) would be used post-1945 to build the ‘Standard Model’ of particle physics using ‘symmetry groups’. In a serendipitous 1972 meeting in the Institute of Advanced Study cafeteria, it was discovered that the distribution of prime numbers has a still-mysterious connection with the energy levels of particles. G.H. Hardy famously wrote, in ‘A Mathematician’s Apology’ which influenced many future mathematicians, that the field of number theory was happily ‘useless’ and did not contribute to ‘any warlike purpose’; even as he wrote the words, it was secretly being applied to cryptography and it now forms the basis of secure electronic communications among other things. Perhaps another example will be the ‘Langlands Program’ in pure mathematics which was developed in the 1960’s and work on it is now funded by DARPA (the famous military technology developer) in the hope of practical applications.

Mathematicians invent (or discover?) concepts by abstraction and then discover connections between concepts.* Nature operates with universal laws and displays symmetry and regularity as well as irregularity and randomness.

‘What do we mean by “understanding” something? We can imagine that this complicated array of moving things which constitutes “the world” is something like a great chess game being played by the gods, and we are observers of the game. We do not know what the rules of the game are; all we are allowed to do is to watch the playing. Of course, if we watch long enough, we may eventually catch on to a few of the rules. The rules of the game are what we mean by fundamental physics. Even if we knew every rule, however, we might not be able to understand why a particular move is made in the game, merely because it is too complicated and our minds are limited. If you play chess you must know that it is easy to learn all the rules, and yet it is often very hard to select the best move or to understand why a player moves as he does. So it is in nature, only much more so; but we may be able at least to find all the rules. Actually, we do not have all the rules now. (Every once in a while something like castling is going on that we still do not understand.) Aside from not knowing all of the rules, what we really can explain in terms of those rules is very limited, because almost all situations are so enormously complicated that we cannot follow the plays of the game using the rules, much less tell what is going to happen next. We must, therefore, limit ourselves to the more basic question of the rules of the game. If we know the rules, we consider that we “understand” the world.’ Richard Feynman.

These physical laws, or rules, use mathematicians’ abstractions.**

There is an amazing connection between mathematicians’ aesthetic sense of ‘beauty’ and their success in finding solutions:

‘It is efficient to look for beautiful solutions first and settle for ugly ones only as a last resort… It is a good rule of thumb that the more beautiful the guess, the more likely it is to survive.’ Timothy Gowers.

‘[S]ciences do not try to explain, they hardly even try to interpret, they mainly make models. By a model is meant a mathematical construct which, with the addition of certain verbal interpretations, describes observed phenomena. The justification of such a mathematical construct is solely and precisely that it is expected to work – that is, correctly to describe phenomena from a reasonably wide area. Furthermore, it must satisfy certain aesthetic criteria – that is, in relation to how much it describes, it must be rather simple… If only relatively little has been explained, one will absolutely insist that it should at least be done by very simple and direct means.’ Von Neumann.

Some of these models allow relatively precise predictions about a particular physical system: for example, Newton’s equations for classical mechanics or the equations for ‘quantum electrodynamics’. Sometimes they are statistical predictions that do not say how a specific event will turn out but what can be expected over a large number of trials and with what degree of confidence: ‘the epistemological value of probability theory is based on the fact that chance phenomena, considered collectively and on a grand scale, create a non-random regularity’ (Kolmogorov). The use of statistical models has touched many fields: ‘Moneyball’ in baseball (the replacement of scouts’ hunches by statistical prediction), predicting wine vintages and ticket sales, dating, procurement decisions, legal judgements, parole decisions and so on.

For example, many natural (e.g. height, IQ) and social (e.g. polling) phenomena follow the statistical theorem called the Central Limit Theorem (CLT) and produce a ‘normal distribution’, or ‘bell curve’. Fields Medallist Terry Tao describes it:

‘Roughly speaking, this theorem asserts that if one takes a statistic that is a combination of many independent and randomly fluctuating components, with no one component having a decisive influence on the whole, then that statistic will be approximately distributed according to a law called the normal distribution (or Gaussian distribution), and more popularly known as the bell curve…

‘The law is universal because it holds regardless of exactly how the individual components fluctuate, or how many components there are (although the accuracy of the law improves when the number of components increases); it can be seen in a staggeringly diverse range of statistics, from the incidence rate of accidents, to the variation of height, weight, or other vital statistics amongst a species, to the financial gains or losses caused by chance, to the velocities of the component particles of a physical system. The size, width, location, and even the units of measurement of the distribution varies from statistic to statistic, but the bell curve shape can be discerned in all cases.

‘This convergence arises not because of any “low-level” or “microscopic” connection between such diverse phenomena as car crashes, human height, trading profits, or stellar velocities, but because in all of these cases the “high-level” or “macroscopic” structure is the same, namely a compound statistic formed from a combination of the small influences of many independent factors. This is the essence of universality: the macroscopic behaviour of a large, complex system can be almost totally independent of its microscopic structure.

‘The universal nature of the central limit theorem is tremendously useful in many industries, allowing them to manage what would otherwise be an intractably complex and chaotic system. With this theorem, insurers can manage the risk of, say, their car insurance policies, without having to know all the complicated details of how car crashes actually occur; astronomers can measure the size and location of distant galaxies, without having to solve the complicated equations of celestial mechanics; electrical engineers can predict the effect of noise and interference on electronic communications, without having to know exactly how this noise was generated; and so forth.’

Many other phenomena (e.g. terrorist attacks, earthquakes, stock market panics) produce a ‘power law’ and trusting to a CLT model of a phenomenon when it actually follows a power law causes trouble, as with the recent financial crisis. When examining phase transitions of materials (e.g the transition from water to ice), the patterns formed by atoms are almost always fractals which appear everywhere from charts of our heartbeats to stock prices to Bach. (Recent work (here) has made breakthroughs in understanding the statistics of phase transitions.)

However, even our best understood mathematical models can quickly become practically overwhelming. Laplace voiced a famous expression of the post-Newton Enlightenment faith in science’s potential to predict.

‘We may regard the present state of the universe as the effect of its past and the cause of its future. An intellect which at a certain moment would know all the forces that animate nature, and all positions of the beings that compose it, if this intellect were vast enough to submit the data to analysis, would condense in a single formula the movements of the greatest bodies of the universe and those of the tiniest atom; for such an intellect nothing would be uncertain and the future just like the past would be present before its eyes… Present events are connected with preceding ones by a tie based upon the evident principle that a thing cannot occur without a cause that produces it… All events, even those which on account of their insignificance do not seem to follow the great laws of nature, are a result of it just as necessarily as the revolutions of the sun.’ Laplace

Newton himself had warned of the potential complexity of calculating more than two interacting bodies.

‘The orbit of any one planet depends on the combined motions of all the planets, not to mention the action of all these on each other. But to consider simultaneously all these causes of motion and to define these motions by exact laws allowing of convenient calculation exceeds, unless I am mistaken, the force of the human intellect.’

It turned out that Newton’s famous gravitational equation cannot be extended to just three bodies without producing ‘deterministic chaos’, so although ‘cosmologists can use universal laws of fluid mechanics to describe the motion of entire galaxies, the motion of a single satellite under the influence of just three gravitational bodies can be far more complicated’ (Tao). Deterministic chaos, a system which is ‘sensitive to initial conditions’, was first articulated by Poincaré as he struggled to solve the ‘three-body problem’, and broke Laplace’s dream of perfect understanding and prediction:

‘If one seeks to visualize the pattern formed by these two [solution] curves and their infinite number of intersections, . . .[their] intersections form a kind of lattice-work, a weave, a chain-link network of infinitely fine mesh; … One will be struck by the complexity of this figure, which I am not even attempting to draw. Nothing can give us a better idea of the intricacy of the three-body problem, and of all the problems of dynamics in general…

‘A very small cause which escapes our notice determines a considerable effect that we cannot fail to see, and then we say that that effect is due to chance. If we knew exactly the laws of nature and the situation of the universe at the initial moment, we could predict exactly the situation of that same universe at a succeeding moment. But even if it were the case that the natural laws had no longer any secret for us, we could still only know the initial situation approximately. If that enabled us to predict the succeeding situation with the same approximation, that is all we require, and we should say that the phenomenon had been predicted, that it is governed by laws. But it is not always so; it may happen that small differences in the initial conditions produce very great ones in the final phenomena. A small error in the former will produce an enormous error in the latter. Prediction becomes impossible, and we have the fortuitous phenomenon.’ (Poincaré, Science and Method, 1913)

Even with systems displaying chaos because of sensitivity to initial conditions, short-term predictions are not hopeless. The best example is weather – the study of which was actually the prompt for Lorenz’s re-discovery of ‘chaos’. Weather forecasts have improved greatly over the past fifty years. For example, 25 years ago forecasts of where a hurricane would hit land in three days time missed by an average of 350 miles; now they miss by about 100 miles. We have bought ourselves an extra 48 hours to evacuate. Is a weather forecast better than it would be by simply a) looking at historical data (climatology), or b) assuming tomorrow will be similar to today (persistence)? Our forecasts are significantly better until about day 9 when forecasts become no better than looking at historical data.

However, chaos means that beyond the short-term, forecasts rapidly break down and usually greater and greater resources are needed to extend the forecasts even just a little further; for example, there has been a huge increase in computer processing applied to weather forecasts since the 1950’s, just to squeeze an accurate forecast out to Day 9. (Cf. Nate Silver’s ‘The signal and the noise‘ for more details.)

‘Even when universal laws do exist, it may still be practically impossible to use them to predict what happens next. For instance, we have universal laws for the motion of fluids, such as the Navier-Stokes equations, and these are certainly used all the time in such tasks as weather prediction, but these equations are so complex and unstable that even with the most powerful computers, we are still unable to accurately predict the weather more than a week or two into the future.’ (Tao)

Between the precision of Newtonian mechanics (with a small number of interacting agents) and the statistics of multi-agent systems (such as thermodynamics and statistical mechanics) ‘there is a substantial middle ground of systems that are too complex for fundamental analysis, but too simple to be universal. Plenty of room, in short, for all the complexities of life as we know it’ (Tao).

Conclusion

In England, less than 10 percent per year leave school with formal training in basics such as ‘normal distributions’ and conditional probability. Less than one percent are well educated in the basics of how the ‘unreasonable effectiveness of mathematics’ provides the language of nature and a foundation for our scientific civilisation. Only a small subset of that <1% then study trans-disciplinary issues concerning complex systems. This number has approximately zero overlap with powerful decision-makers.

Generally, they are badly (or narrowly) educated and trained. Even elite universities offer courses such as PPE that are thought to prepare future political decision-makers but are clearly inadequate and in some ways damaging, giving people like Cameron and Balls false confidence in 1) the value of their acquired bluffing skills and 2) the scientific basis of modern economics’ forecasts. Powerful decision-makers also usually operate in institutions that have vastly more ambitious formal goals than the dysfunctional management could possibly achieve, and which generally select for the worst aspects of chimp politics and against those skills seen in rare successful organisations (e.g the ability to simplify, focus, and admit errors). Most politicians, officials, and advisers operate with fragments of philosophy, little knowledge of maths or science (few MPs can answer even simple probability questions yet most are confident in their judgement), and little experience in well-managed complex organisations. The skills, and approach to problems, of our best mathematicians, scientists, and entrepreneurs are almost totally shut out of vital decisions.

These issues are connected to the failure of political elites to get big decisions right since the 1860s, as I discussed in The Hollow Men. In Part II next week, I will discuss some of the issues about how Whitehall works that cause so many problems and what can be done to improve this situation. In Part II of this blog, I will explore some more of the science of prediction. But I’d prefer you to look at my essay, from which most of this is taken…

* This happens in social sciences too. E.g. Brouwer’s fixed-point theorem in topology was first applied to ‘equilibrium’ in economics by von Neumann (1930’s), and this approach was copied by Arrow and Debreu in their 1954 paper that laid the foundation for modern ‘general equilibrium theory’ in economics.

** Einstein asked, ‘How is it possible that mathematics, a product of human thought that is independent of experience, fits so excellently the objects of physical reality?’ ‘Is mathematics invented or discovered?’, Tim Gowers (Polkinghorne, 2011). Hilbert, Cantor and Einstein thought it is invented (formalism). Gödel thought it is discovered (Platonism). For a non-specialist summary of many issues concerning maths and prediction, cf. a talk by Fields Medallist Terry Tao. Wigner answered Einstein in a famous paper, ‘The Unreasonable Effectiveness of Mathematics in the Natural Sciences’ (1960).

April 8, 2014 by dominiccummings

Some interesting things I’ve read recently (March)

This PDF has links to some interesting things I’ve read recently.

They are organised in the same framework as my essay.

Education stuff mostly in Section 6.

March 10, 2014 by dominiccummings

Some interesting things I’ve read recently

This PDF summarises some interesting pieces I’ve read recently, up to the end of February 2014.

They are organised in the same framework as my essay.

November 19, 2013 by dominiccummings

An essay: ‘Some thoughts on education and political priorities’

On the evening of Friday 11 October 2013, the Guardian published a draft essay of mine – ‘Some thoughts on education and political priorities’.

This page has links to various things about it.

This twitter feed has links on themes in the essay: @OdysseanProject

Dominic Cummings's Blog

‘Odyssean’ education

Times op-ed: What Is To Be Done? An answer to Dean Acheson’s famous quip

Complexity, ‘fog and moonlight’, prediction, and politics II: controlled skids and immune systems (UPDATED)

‘Standin’ by the window, where the light is strong’: de-extinction, machine intelligence, the search for extra-solar life, autonomous drone swarms bombing Parliament, genetics & IQ, science & politics, and much more @ SciFoo 2014

Complexity, ‘fog and moonlight’, prediction, and politics I

Some interesting things I’ve read recently (March)

Some interesting things I’ve read recently

An essay: ‘Some thoughts on education and political priorities’

Share this:

Share this:

Share this:

Share this:

Share this:

Share this:

Share this: