OK, so the other week everyone who was anyone was banging on about the AlphaGo programme from The Google’s DeepMind that can, so far, beat humans at Go. Baruch’s understanding is that this was a different type of AI than we’ve seen with Deep Blue and Watson kicking human butt at chess and Jeopardy. Both of those deployed the main trump card that computers have against human minds — the ability to scan, prioritise, sort and rank huge databases extremely rapidly to come up with statistically likely solutions to well defined questions. All they needed to defeat the humans were mega amounts of computer processing, clever coders, a ludic fallacy and a large, cold room to keep the servers in.
DeepMind, so far as I am told, uses something more akin to analogy in its core processing, with a feedback loop that keeps the focus on a narrowly defined and therefore manageable number of moves. Words Baruch pretends to understand but probably doesn’t, like “neural networks” and “deep learning”have been bandied about. To play Go well, you clearly have to be able to program smart. That said, AlphaGo doesn’t seem terribly parsimonious in its use of processing grunt: a “mere” 2000 CPUs, as well as 280 high end GPUs using what we experts* call parallel processing, so perhaps it’s still just the typical thing of throwing Moore’s Law at something until you have enough computing power to make it work eventually. But it probably isn’t, in which case chapeau, DeepMind.
So overall, it’s mildly interesting, and at a more mercenary level, grist to Baruch’s mill. If the advances made by DeepMind bring us one step closer to the massive disruption digital automation is going to wreak on business and society, then I’m going to be quids in. My current project is about providing investors with an opportunity to take advantage of those transformations, so this can only encourage more people to want to invest when my idea finally goes live. Yay!
We all need to hedge against the coming AI whirlwind, and this is Baruch’s plan. However, his biggest worry is resistance is futile — that even his job as a fund manager is going be replaceable by machines, like all of yours will be. So imagine his horror when he saw this (HT Alphaville) in Wired from back in January.
LAST WEEK, BEN Goertzel and his company, Aidyia, turned on a hedge fund that makes all stock trades using artificial intelligence—no human intervention required.
Oh bollocks, he thought.
But at times like this I think of Douglas Adams, and have learned not to panic.
The fact is, a lot of this AI based digital transformation is actually crap. Enough of it isn’t for it to still matter, very very much. Like I imply above, I don’t know enough to judge whether this is true of AlphaGo, but after a closer reading of the Wired article I began to feel a little less worried about Adiya’s particular bot.
For starters, the exciting-sounding AI technique of “evolutionary computation” struck me as suspiciously akin to the Lucky Fool metaphor, outlined in Nicholas Taleb’s excellent Fooled By Randomness, albeit one expanded from 10,000 fools (10 kilofools) into hundreds of Tera-, Exa- and Zettafools:
it creates a large and random collection of digital stock traders and tests their performance on historical stock data. After picking the best performers, it then uses their “genes” to create a new set of superior traders. And the process repeats. Eventually, the system homes in on a digital trader that can successfully operate on its own. “Over thousands of generations, trillions and trillions of ‘beings’ compete and thrive or die,” Blondeau says, “and eventually, you get a population of smart traders you can actually deploy.”
Of course this sounds great, except there is nothing here to say that the “population of smart traders” will have any predictive value whatsoever. For even if the trillions of traders’ outcomes were decided randomly, i.e. not one of them had any “skill”, a population of winning traders is exactly what you would achieve anyway. It’s maths, innit. Perhaps it’s smarter than that, but as described here Adiya’s methodology strikes me as just an extremely expensive and futile exercise in backtesting at massive scale.
Baruch is not a fan of backtesting.
As for the rest, “deep learning”, like “evolutionary computing”, sounds flipping amazing, but in the end again you are throwing millions of dollars and teraflops at something to get a result which is only a bit better than what a puny fleshbag human can achieve at a fraction of the cost. Currently. But it’s not always clear either if that edge can scale with Moore’s Law, as computers can and will be gamed, if not by the Humans** then certainly by other computers.
This is the problem with analysing liquid markets, as opposed to analysing traffic patterns, playing chess or Go, or predicting when cows are ready for oestrus, all laudable uses of Big Data and AI. Financial market analysis eventually changes the outcome under analysis in a feedback loop that invalidates the original analysis. If you see what I mean. Baruch’s a fan (to a certain extent) of Quants, and like e.g. plate spinners has great respect for them even if he lacks the interest and ability to join in. But I’d bet the guys at Renaissance and DE Shaw don’t let the bots work by themselves; I reckon they’re “centaurs” or cyborgs, where the humans stay in charge to program or monitor the bots and to intervene when the strategies become stale, i.e. when the market (or other bots) adapt to take advantage of the previous dominant bot strategy or combination thereof. These guys don’t pay millions to Russian maths prodigies for nothing.
Imagine a stockmarket where the Quants, like the Indexers seem to want to do, had finally won and defeated the Humans. It’d be awful. There’d be no yardstick of economic value. Prices would be decided not by the ability of companies to pay dividend streams in the future but by algos of increasing sophistication and timeframes gaming each other. I’d love to see what would happen! Maybe prices would simply flatline until some newly effective AI system popped up, and suddenly zoom off in a previously unknowable direction before that AI was gamed and its edge blunted. Maybe there’d be an “ultima-bot” which would instantly analyse the edge of the new bot and work out a profitable strategy to milk and ultimately defeat it. There’d be no value in the stockmarket as a source of finance for the economy. It would be zero sum. Quants would probably have to end up inventing increasingly expensive sophisticated bots which acted like humans (and take turns losing money) in order for it all to work. Or companies would have to invent a “humans only” stockmarket where they could raise money and their valuations can reflect, their, you know, actual value.
Because ultimately that’s what stock markets are for, and quants like Adiya and other bots don’t have anything to say about it. Like indexers, the bots are free riders, potential threats to society, parasites that need to be kept under control lest they destroy the host organism. Breathless journo clickbait (or the 4th most successful folk parody band from New Zealand) aside, they are not going to take over the stockmarket. Not to say they won’t take over lots of things, but there’s still enough messiness around to make sure that the humans still have to be involved somewhere with a hand on the off switch. We may need fewer of the humans, is all.
*Baruch is proud to announce he took an internet course in coding recently on EDX. OK I didn’t actually finish it. I gave up after 2 weeks because it was so damn hard and (I thought) badly taught, and was taking up time I needed instead to get a new gig. What Baruch did relearn, however, is that computers are dreadfully, unimaginably stupid and evil-tempered things, and like a 12 year old girl cleaning up the kitchen need to be told exactly what to do and in what order or they throw their arms up in the air, sigh exaggeratedly and storm off in the digital equivalent of a huff. Respect for coders.
**My friend Julien can regularly beat the AI at Civilization V at DEITY level, something supposed to be near impossible. He’s worked it out over months, something to do with playing the Persians, building an early Academy and going for science victory. Conversely, Baruch’s stats on World of Tanks have taken a turn for the worse now he’s in the higher tiers. That’s playing humans for you.