Discover more from Dan Davies - "Back of Mind"
the statistical heart of the system
planned economies and different styles
This is now an occasional series of pieces related to what did and didn’t go on in Allende’s Chile when Stafford Beer was involved in the big experiment in running an economy on cybernetic principles. One thing I’ve tried to emphasise so far is that CYBERSYN really, really wasn’t a “planned economy”. This is a bit of a difficult one to make, though, because it happens to be the case that the Soviet Union used to call its system “economic cybernetics”.
The reason why they used this nomenclature is basically that Russians also read “Cybernetics” by Norbert Wiener, which was an absolutely huge best seller in its day – sort of “Selfish Gene” levels of sales, which is quite extraordinary given that it has several chapters in which there are consecutive pages of nothing but equations. Its influence persists in really funny little corners to this day – it spawned a rip-off self-help book called “Psycho-Cybernetics” which is the basis of the Tony Robbins Success Formula, and L Ron Hubbard’s “Dianetics” was the subject of a quite cross correspondence with Wiener who absolutely forbade him to imply any endorsement.
for ideas behind the news, and snippets of my forthcoming book on industrialised decision making and the end of accountability.
But anyway, Soviet cybernetics tended to be based on big linear programming models, as described in Francis Spufford’s book “Red Plenty”. Given the computing power of the day, these models absolutely weren’t up to the task – they couldn’t handle large enough input-output matrices, so things had to be aggregated far too much, and they took ages to deliver results, meaning that everything was out of date before it could be used to support decisions. But the underlying maths is that of optimisation – you have a bunch of inputs, you want a bunch of outputs. You specify a bunch of equations describing the relationship between them, then you follow Kantorovich’s method to get the optimal solutions to those equations.
Quite interestingly, some people today (like Paul Cockshott) think that poor old Kantorovich might not have been given a fair go, and that fully automated democratic planning could be possible today. Partly because of much faster computers, but mainly because of better algorithms – most inputs aren’t involved in producing most outputs, and we have much better techniques than were available in the 1940s for dealing with matrices where the majority of items are zeroes. I am very much not in a position judge whether this is accurate or not – since companies the size of WalMart do in fact run computer models to forecast demand and optimise supply, I don’t think it can be ruled out.
But in any case, the piece of computer software that Stafford Beer commissioned for Chile (Cyberstride) was not at all like that. It was basically a Bayesian classifier. It took time series data, mainly for production outputs, and decided whether the latest data points were most likely to represent:
i) noise, in a basically stationary series (or around a stationary trend)
ii) an outlier that would be reversed
iii) a new linear trend or change in slope of existing trend
iv) a new exponential trend
v) a level shift
Based on this classification, it would raise flags with the basic semantic content “something’s up at t’ mill”, for someone to investigate more fully. (Note here that the Chilean system was trying to respect the problem – rather than trying to use the data itself to guess what was happening, it would trigger analysis by a person or team who would start by establishing ground truth, and which had a better chance of delivering sufficient bandwidth to handle the complexity of what would happen).
Cyberstride was one component of the Chilean economic management project – the other two big components were meant to be a) the massive communications network needed to deliver the input to Cyberstride and b) lots and lots of teams of managers and analysts responsible for deciding which time series ought to be collected.
The other thing which Beer wanted to be designed into Cyberstride is that it was meant to be usable at any level of management. There was no realistic prospect of computerising the whole Chilean economy (there were about fifty computers in the country, most of them quite old, and there was both a shortage of foreign exchange and a soft-sanctions regime on technology imports from the USA). But the idea of using data to throw up exceptions and focus management effort wasn’t intrinsically central – it was meant to be used as the fundamental method at every level of the economy.