In the near future we may face a potent new enemy. Some will say it is not really an enemy, but a powerful foreigner with whom we should do business. This entity will have goals that are unlike our own or simply opaque. Yet people will argue that it can be reasoned with. People will say it’s not a direct threat yet, but if it ever is we’ll simply deal with it then.
I am talking about a speculative, future artificial general intelligence (AGI). It would be powerful, perhaps more powerful than us, but we hope it would be benevolent. But I could also be talking about a new dictator and the opening paragraph contains actual responses to Hitler in the 1930s.
Hold that thought, because today I speak of our overreliance on historical examples.
The chronicle of history is full of people being utterly shocked by unprecedented events. Human history prior to the chronicles was probably more stable. But with the shift to agriculture, permanent settlements, stockpiles, and dynasties — hence the need for written chronicles — history’s undulations could displace or destroy whole populations. At your time of death, the world could be radically unlike it was when you were born. And nowadays, since the greater acceleration of modernity, it’s assumed that the world will transform in many ways in just one generation.
I suspect that this increasingly rapid change makes history less and less relevant for understanding the present. It now takes less time for conditions to change, and therefore less time for one epoch to cease resembling the previous one. It’s unsatisfying, but the past is not a firm guide to the increasingly different future. History isn’t what it used to be.
In philosophy, history’s inadequacy as a guide to the future is known as the problem of induction. Induction (or inductive reasoning) is the method of observing a bunch of examples of something and then finding a pattern among them. For some, this is what science is: find a pattern in the existing data.
Pattern-finders need to beware. Just because some regularity has been observed in the past it does not follow that it will continue in the future. Assuming it does can have catastrophic consequences.1 Nassim Nicholas Taleb vividly dramatises it as “the turkey problem”. The turkey notices that 1000 days in a row the farmer has fed it every morning. It has total confidence, based on past experience, that it will be fed tomorrow as well. Except that tomorrow is Thanks Giving in America, so instead of feeding the turkey, the farmer cuts its throat.
I like this line from the investor Morgan Housel:
“The correct lesson to learn from surprises is that the world is surprising.
This is the only lesson we can draw from history too. Is it contradictory to say that what we learn from the past is that we cannot learn from the past?
Meh. Here’s a good way to think about it. If the world had been surprising just once, that would tell us that we can’t totally rely on past regularities (induction as a universal law would be falsified or discredited).
And even if the past had never failed yet, if we lived in a world where our careful study of past trends thus far worked to predict the future, we should still be able to imagine a dreadful asymmetry between possible worlds. Consider these two worlds:
Kaos: a world of randomness and irregularity that is impossible to predict simply from spotting patterns. The data are all over the place. No one knows what kind of world it will be tomorrow.
Kosmos: a world of perfect regularity that yields to inductive inferences about patterns and laws. You notice a trend, it continues into the future. You bet everything on tomorrow being like today.
Here’s the asymmetry.
The first world, Kaos, could go through some random fluctuations that, for a time, resemble an orderly, predictable world, before reverting to more haphazard ways. This would lure observers — those who just happened to be alive and observing during the temporary spell of regularity — into a false hope that they can learn from the past and extrapolate to the future. They would believe they are living in Kosmos.
But the reverse cannot happen. Kosmos cannot mimic Kaos. As soon as it is disorderly for a while, it’s no longer orderly.2
History is so diverse it can support any argument
Guys, we live in Kaos. But for large tracts of early human history it might have looked like Kosmos.3 We can pretend we live in Kosmos by cherry-picking from history only the examples that support our case. But we should be suspicious of how easily people can marshal historical examples to provide narratives for any argument.
Here are some examples of historical examples for arguing about AI:
In support of how bans on technologies can work and are adhered to: human cloning.
In support of the inevitability of new technologies being adopted, despite regulatory lag and opposition from incumbents: Uber.
New technology dealing with safety concerns: commercial flight.
Technology can be more trouble than it was worth: the agricultural revolution (which apparently made everyone more malnourished, overworked, diseased, etc.)
Unintended consequences were bad: CFCs.
Unintended consequences were good: GPUs were designed for gaming but can now power AI.4
Technology is overhyped: fusion power.
Technology is underappreciated: washing machines.
Scientists are foolhardy: the development of the atomic bomb.
Cooler heads prevail: the Cuban Missile Crisis.5
(I haven’t spent much time on this, others can summon more/better examples. Indeed check out this great post from dynomight whose Substack is cool.)
AI is new. People want an analogy from history that suits their intuitions. History is broad and dense. For those seeking polemical examples to spice up their arguments, history is a grand bazaar of whatever you need.
If you don’t want to spend time looking, just use the brief history of AI itself. It was overhyped in the 1950s, 1970s, 1990s, ergo it must be overhyped now. This is the kind of inductive fallacy that even GPT-4 can diagnose.
I think history is a source of negative information. You can always find an example to prove your point. But the fact that you can find counterexamples as well should prove the larger point: there are no simple historical narratives, no immutable laws, and a single vivid example is not a good argument.6
Thinking about AI, I propose that historical precedent is irrelevant. Let’s say you wanted to ban some version of AI. It doesn’t matter if there are or aren’t examples of moratoria on technologies in the past. We either can or cannot do that in the future depending on our actions. And it doesn’t matter that most technologies have typically not destroyed the world or vastly improved it. The current technology might be unique. Tbh, it is.
What touchstones from history should we choose?
History provides flashy case studies and fodder for moral fables. And after having just written a thousand words against using history as a guide, or even trying to interpret history, I now return to my colourful and tendentious historical analogy for AI: Churchill versus Chamberlain.
What I’m not saying with this analogy: Churchill was proved right in 1939 with his alarmism over Hitler, therefore alarmism is always appropriate.
Nor are the details of pre-War Europe all that similar to our details. And the AGI I’m talking about doesn’t even exist yet.
I am saying: Chamberlain was doing wishful thinking, Churchill wasn’t. In that regard, we can try to emulate Churchill.7 Not because our circumstances are similar, forming a pattern with the past, but because the dangers of wishful thinking in high-stakes situations are grave. And there is one other feature shared with our circumstances: we’re talking about giving more power to something that you already struggle to control and whose interests may not align with yours.
Think how tempting it was for leaders around Europe. Maybe Hitler will stop with his pesky encroachments if some concessions are made. A war was expensive, destructive, unthinkable. It would be better for Britain, certainly, if Germany wasn’t confronted.
The definition of wishful thinking is believing something is true because you want it to be true. It’s enticing, especially when you have no control over events; if you can’t control the outcome you can at least control the way you think about it, so might as well hope for the best and sleep easy. But when you can control events, wishful thinking is an unforgivable abdication of responsibility.
Perhaps Chamberlain actually thought conflict was impossible so he defaulted to believing what he wanted to be true: that Hitler’s motives could be divined, his next actions predicted, his desire sated. Some degree of “alignment” with the Nazis could be achieved by ceding Czechoslovakia. The pipedream of a benevolent dictator.
Churchill thought the situation was controllable but might not be for much longer. Some showdown with Hitler was inevitable. Better to have it sooner rather than later by which time he might be too powerful.8
The appeasers think that if we just negotiate with the AGI we could fix all our global issues. Surely any general intelligence worth the name, that has been endowed by its creator (us) with the ability to reason, will be rational and interested in cooperation and will share our horror of violence and war. That’s right, us. We who include Chamberlain, Churchill, and Hitler. Just get those guardrails in place, solve the alignment problem, and relax under the benevolent dictatorship of a machine more powerful than ourselves.
And of course, it’s possible. It must be. Even human dictators have occasionally been benevolent. History has examples of everything if you’re selective.
But what’s the problem with benevolent dictators? That you have no control over them. If they turn bad, you’re fucked. If they die and their successor is bad, you’re fucked. Even when a dictator really is benevolent, it’s an inherently unsustainable existence. I hope this seems obvious, but there are always people everywhere who propose a benevolent dictator to solve their problems. For the love of Elon, even when a dictator is openly malevolent, like Hitler, a subset of people still advocate them to solve all their problems.
Ceding control to an AGI in the future is inherently unsustainable (for us — it may be sustainable for the AGI) even if it initially or temporarily seems stable. Kaos, Kosmos, etc.
But wherever this journey leads, and however long it takes to develop AGI, there will be people advocating it at every stop, no matter how salient the risks are.
Yudkowsky has some peccadilloes. He comes close to having an ideology of infinity: where something is either infinitely good or infinitely bad, justifying any action to bring it about or avoid it. And he’s apparently not so rational that he’ll make a minor and good-for-everyone-involved fashion change, even if it improves his chances of saving the world.
But Hell’s-Easier Yudkowsky is not a Chamberlain; he’s a Churchill. According to his best estimates he sees darkness ahead. Does he settle for the opiate of wishful thinking? No. He bays from the political wilderness. Steadfast, he vows to fight on, to fight the AI dreamers on social media, to fight the alignment naysayers in podcast interviews, to fight on in obscure reports published by MIRI, and never to surrender.
I close, pointedly, with Churchill on appeasement:
“And do not suppose that this is the end. This is only the beginning of the reckoning. This is only the first sip, the first foretaste of a bitter cup which will be proffered to us year by year unless by a supreme recovery of moral health and martial vigour, we arise again and take our stand for freedom as in the olden time.”
Another irony is that the naive approaches to AI are themselves based in mere pattern detection. The more powerful kinds of intelligence and the more respectable forms of science try to go beyond this. They interrogate the patterns and counter-patterns to account for what produces the pattern and then try to intervene in the system to cause new patterns to happen.
This mirrors Taleb’s distinction between mediocristan and extremistan.
Jordan Peterson fanboys (any fangirls?) will bristle. “Aren’t all our myths about dealing with chaos?” Ironically, myth-heads like Peterson are more apt than anyone to find patterns across all human cultures and histories. But I’ll say this once: there is no settled theory or explanation of the function of any one myth, let alone myth in general. And if the deep past with its hundreds of generations of continuous culture and static technology was a fight to keep chaos at bay, then how the fuck would cultural technologies calibrated to then be of any use now?
It is actually harder for me to think of positive unintended effects. On the one hand I think most technologies get repurposed for uses their creators couldn’t envision: positive unintended consequences. But there might also be a tilt to disorder so that it’s easier to break systems than enhance them, so unintended consequences are more likely. I don’t know!
This is a bitter joke.
When I worked at the Institute for Advanced Studies in the Humanities most of my colleagues were historians. Folks, I respect historians. In my even-more-cynical moments I feel like they’re the only true scholars; them and mathematicians. Real academic historians are constantly trying to reassess the past which is in no way settled or predictable. This should humble anyone foolhardy enough to predict the future. It also made me realise I knew nothing about anything, because everything has has a history and it is always more complicated and multifarious than the little cartoon outline in one’s head.
And Churchill was a colonial shitbag in multiple ways, Indians in particular are right to vilify him, Gallipoli, and so on. I don’t care about Churchill’s larger legacy or him as a person; he’s just a great example of a realist at the right time and place in 1939.
Historians will know that these caricatures of Chamberlain and Churchill are inaccurate. Churchill was sounding off about Hitler, but oftentimes in a politically self-serving way; Chamberlain wasn’t the feckless Hitler-simp I’ve portrayed above; etc.