I.
Imagine a distant planet full of eyeless animals. Evolving eyes is hard: they need to evolve Eye Part 1, then Eye Part 2, then Eye Part 3, in that order. Each of these requires a separate series of rare mutations.
Here on Earth, scientists believe each of these mutations must have had its own benefits – in the land of the blind, the man with only Eye Part 1 is king. But on this hypothetical alien planet, there is no such luck. You need all three Eye Parts or they’re useless. Worse, each Eye Part is metabolically costly; the animal needs to eat 1% more food per Eye Part it has. An animal with a full eye would be much more fit than anything else around, but an animal with only one or two Eye Parts will be at a small disadvantage.
So these animals will only evolve eyes in conditions of relatively weak evolutionary pressure. In a world of intense and perfect competition, where the fittest animal always survives to reproduce and the least fit always dies, the animal with Eye Part 1 will always die – it’s less fit than its fully-eyeless peers. The weaker the competition, and the more randomness dominates over survival-of-the-fittest, the more likely an animal with Eye Part 1 can survive and reproduce long enough to eventually produce a descendant with Eye Part 2, and so on.
There are lots of ways to decrease evolutionary pressure. Maybe natural disasters often decimate the population, dozens of generations are spent recolonizing empty land, and during this period there’s more than enough for everyone and nobody has to compete. Maybe there are frequent whalefalls, and any animal nearby has hit the evolutionary jackpot and will have thousands of descendants. Maybe the population is isolated in little islands and mountain valleys, and one gene or another can reach fixation in a population totally by chance. It doesn’t matter exactly how it happens, it matters that evolutionary pressure is low.
The branch of evolutionary science that deals with this kind of situation is called “adaptive fitness landscapes”. Landscapes really are a great metaphor – consider somewhere like this:
You pour out a bucket of water. Water “flows downhill”, so it’s tempting to say something like “water wants to be at the lowest point possible”. But that’s not quite right. The lowest point possible is the pit, and water won’t go there. It will just sit in the little puddle forever, because it would have to go up the tiny little hillock in order to get to the pit, and water can’t flow uphill. Using normal human logic, we feel tempted to say something like “Come on! The hillock is so tiny, and that pit is so deep, just make a single little exception to your ‘always flow downhill’ policy and you could do so much better for yourself!” But water stubbornly refuses to listen.
Under conditions of perfectly intense competition, evolution works the same way. We imagine a multidimensional evolutionary “landscape” where lower ground represents higher fitness. In this perfectly intense competition, organisms can go from lower to higher fitness, but never vice versa. As with water, the tiniest hillock will leave their potential forever unrealized.
Under more relaxed competition, evolution only tends probabilistically to flow downhill. Every so often, it will flow uphill; the smaller the hillock, the more likely evolution will surmount it. Given enough time, it’s guaranteed to reach the deepest pit and mostly stay there.
Take a moment to be properly amazed by this. It sounds like something out of the Tao Te Ching. An animal with eyes has very high evolutionary fitness. It will win at all its evolutionary competitions. So in order to produce the highest-fitness animal, we need to – select for fitness less hard? In order to produce an animal that wins competitions, we need to stop optimizing for winning competitions?
This doesn’t mean that less competition is always good. An evolutionary environment with no competition won’t evolve eyes either; a few individuals might randomly drift into having eyes, but they won’t catch on. In order to optimize the species as much as possible as fast as possible, you need the right balance, somewhere in the middle between total competition and total absence of competition.
In the esoteric teachings, total competition is called Moloch, and total absence of competition is called Slack. Slack (thanks to Zvi Mowshowitz for the term and concept) gets short shrift. If you think of it as “some people try to win competitions, other people don’t care about winning competitions and slack off and go to the beach”, you’re misunderstanding it. Think of slack as a paradox – the Taoist art of winning competitions by not trying too hard at them. Moloch and Slack are opposites and complements, like yin and yang. Neither is stronger than the other, but their interplay creates the ten thousand things.
II.
Before we discuss slack further, a digression on group selection.
Some people would expect this discussion to be quick, since group selection doesn’t exist. These people understand it as evolution acting for the good of a species. It’s a tempting way to think, because evolution usually eventually makes species stronger and more fit, and sometimes we colloquially round that off to evolution targeting a species’ greater good. But inevitably we find evolution is awful and does absolutely nothing of the sort.
Imagine an alien planet that gets hit with a solar flare once an eon, killing all unshielded animals. Sometimes unshielded animals spontaneously mutate to shielded, and vice versa. Shielded animals are completely immune to solar flares, but have 1% higher metabolic costs. What happens? If you predicted “magnetic shielding reaches fixation and all animals get it”, you’ve fallen into the group selection trap. The unshielded animals outcompete the shielded ones during the long inter-flare period, driving their population down to zero (though a few new shielded ones arise every generation through spontaneous mutations). When the flare comes, only the few spontaneous mutants survive. They breed a new entirely-shielded population, until a few unshielded animals arise through spontaneous mutation. The unshielded outcompete the shielded ones again, and by the time of the next solar flare, the population is 100% unshielded again and they all die. If the animals are lucky, there will always be enough spontaneously-mutated shielded animals to create a post-flare breeding population; if they are unlucky, the flare will hit at a time with unusually few such mutants, and the species will go extinct.
An Evolution Czar concerned with the good of the species would just declare that all animals should be shielded and solve the problem. In the absence of such a Czar, these animals will just keep dying in solar-flare-induced mass extinctions forever, even though there is an easy solution with only 1% metabolic cost.
A less dramatic version of the same problem happens here on Earth. Every so often predators (let’s say foxes) reproduce too quickly and outstrip the available supply of prey (let’s say rabbits). There is a brief period of starvation as foxes can’t find any more rabbits and die en masse. This usually ends with a boom-bust cycle: after most foxes die, the rabbits (who reproduce very quickly and are now free of predation) have a population boom; now there are rabbits everywhere. Eventually the foxes catch up, eat all the new rabbits, and the cycle repeats again. It’s a waste of resources for foxkind to spend so much of time and its energy breeding a huge population of foxes that will inevitably collapse a generation later; an Evolution Czar concerned with the common good would have foxes limit their breeding at a sustainable level. But since individual foxes that breed excessively are more likely to have their genes represented in the next generation than foxes that breed at a sustainable level, we end up with foxes that breed excessively, and the cycle continues.
(but humans are too smart to fall for this one, right?)
Some scientists tried to create group selection under laboratory conditions. They divided some insects into subpopulations, then killed off any subpopulation whose numbers got too high, and “promoted” any subpopulation that kept its numbers low to better conditions. They hoped the insects would evolve to naturally limit their family size in order to keep their subpopulation alive. Instead, the insects became cannibals: they ate other insects’ children so they could have more of their own without the total population going up. In retrospect, this makes perfect sense; an insect with the behavioral program “have many children, and also kill other insects’ children” will have its genes better represented in the next generation than an insect with the program “have few children”.
But sometimes evolution appears to solve group selection problems. What about multicellular life? Stick some cells together in a resource-plentiful environment, and they’ll naturally do the evolutionary competition thing of eating resources as quickly as possible to churn out as many copies of themselves as possible. If you were expecting these cells to form a unitary organism where individual cells do things like become heart cells and just stay in place beating rhythmically, you would call the expected normal behavior “cancer” and be against it. Your opposition would be on firm group selectionist grounds: if any cell becomes cancer, it and its descendants will eventually overwhelm everything, and the organism (including all cells within it, including the cancer cells) will die. So for the good of the group, none of the cells should become cancerous.
The first step in evolution’s solution is giving all cells the same genome; this mostly eliminates the need to compete to give their genes to the next generation. But this solution isn’t perfect; cells can get mutations in the normal course of dividing and doing bodily functions. So it employs a host of other tricks: genetic programs telling cells to self-destruct if they get too cancer-adjacent, an immune system that hunts down and destroys cancer cells, or growing old and dying (this last one isn’t usually thought of as a “trick”, but it absolutely is: if you arrange for a cell line to lose a little information during each mitosis, so that it degrades to the point of gobbledygook after X divisions, this means cancer cells that divide constantly will die very quickly, but normal cells dividing on an approved schedules will last for decades).
Why can evolution “develop tricks” to prevent cancer, but not to prevent foxes from overbreeding, or aliens from losing their solar flare shields? Group selection works when the group itself has a shared genetic code (or other analogous ruleset) that can evolve. It doesn’t work if you expect it to directly change the genetic code of each individual to cooperate more.
When we think of cancer, we are at risk of conflating two genetic codes: the shared genetic code of the multicellular organism, and the genetic code of each cell within the organism. Usually (when there are no mutations in cell divisions) these are the same. Once individual cells within the organism start mutating, they become different. Evolution will select for cancer in changes to individual cells’ genomes over an organism’s lifetime, but select against it in changes to the overarching genome over the lifetime of the species (ie you should expect all the genes you inherited from your parents to be selected against cancer, and all the mutations in individual cells you’ve gotten since then to be selected for cancer).
The fox population has no equivalent of the overarching genome; there is no set of rules that govern the behavior of every fox. So foxes can’t undergo group selection to prevent overpopulation (there are some more complicated dynamics that might still be able to rescue the foxes in some situations, but they’re not relevant to the simple model we’re looking at).
In other words, group selection can happen in a two-layer hierarchy of nested evolutionary systems when the outer system (eg multicellular humans) includes rules that the inner system (eg human cells) have to follow, and where the fitness of the evolving-entities in the outer system depends on some characteristics of the evolving-entities in the inner system (eg humans are higher-fitness if their cells do not become cancerous). The evolution of the outer layer includes evolution over rulesets, and eventually evolves good strong rulesets that tell the inner-layer evolving entities how to behave, which can include group selection (eg humans evolve a genetic code that includes a rule “individual cells inside of me should not get cancer” and mechanisms for enforcing this rule).
You can find these kinds of two-layer evolutionary systems everywhere. For example, “cultural evolution” is a two-layer evolutionary system. In the hypothetical state of nature, there’s unrestricted competition – people steal from and murder each other, and only the strongest survive. After they form groups, the groups compete with each other, and groups that develop rulesets that prevent theft and murder (eg legal codes, religions, mores) tend to win those competitions. Once again, the outer layer (competition between cultures) evolves groups that successfully constrains the inner layer (competition between individuals). Species don’t have a czar who restrains internal competition in the interest of keeping the group strong, but some human cultures do (eg Russia).
Or what about market economics? The outer layer is companies, the inner layer is individuals. Maybe the individuals are workers – each worker would selfishly be best off if they spent the day watching YouTube videos and pushed the hard work onto someone else. Or maybe they’re executives – each individual executive would selfishly be best off if they spent their energy on office politics, trying to flatter and network with whoever was most likely to promote them. But if all the employees loaf off and all the executives focus on office politics, the company won’t make products, and competitors will eat their lunch. So someone – maybe the founder/CEO – comes up with a ruleset to incentivize good work, probably some kind of performance review system where people who do good work get promoted and people who do bad work get fired. The outer-layer competition between companies will select for corporations with the best rulesets; over time, companies’ internal politics should get better at promoting the kind of cooperation necessary to succeed.
How do these systems replicate multicellular life’s success without being literal entities with literal DNA having literal sex? They all involve a shared ruleset and a way of punishing rulebreakers which make it in each individual’s short-term interest to follow the ruleset that leads to long-term success. Countries can do that (follow the law or we’ll jail you), companies can do that (follow our policies or we’ll fire you), even multicellular life can sort of do that (don’t become cancer, or immune cells will kill you). When there’s nothing like that (like the overly-fast-breeding foxes) evolution fails at group selection problems. When there is something like that, it has a chance. When there’s something like that, and the thing like that is itself evolving (either because it’s encoded in literal DNA, or because it’s encoded in things like company policies that determine whether a company goes out of business or becomes a model for others), then it can reach a point where it solves group selection problems very effectively.
In the esoteric teachings, the inner layer of two-layer evolutionary systems is represented by the Goddess of Cancer, and outer layer by the Goddess of Everything Else. In each part of the poem, the Goddess of Cancer orders the evolving-entities to compete, but the Goddess of Everything Else recasts it as a two-layer competition where cooperation on the internal layer helps win the competition on the external layer. He who has ears to hear, let him listen.
III.
Why the digression? Because slack is a group selection problem. A species that gave itself slack in its evolutionary competition would do better than one that didn’t – for example, the eyeless aliens would evolve eyes and get a big fitness boost. But no individual can unilaterally choose to compete less intensely; if it did, it would be outcompeted and die. So one-layer evolution will fail at this problem the same way it fails all group selection problems, but two-layer systems will have a chance to escape the trap.
The multicellular life example above is a special case where you want 100% coordination and 0% competition. I framed the other examples the same way – countries do best when their citizens avoid all competition and work together for the common good, companies do best when their executives avoid self-aggrandizing office politics and focus on product quality. But as we saw above, some systems do best somewhere in the middle, where there’s some competition but also some slack.
For example, consider a researcher facing their own version of the eyeless aliens’ dilemma. They can keep going with business as normal – publishing trendy but ultimately useless papers that nobody will remember in ten years. Or they can work on Research Program Part 1, which might lead to Research Program Part 2, which might lead to Research Program Part 3, which might lead to a ground-breaking insight. If their jobs are up for review every year, and a year from now the business-as-normal researcher will have five trendy papers, and the groundbreaking-insight researcher will be halfway through Research Program Part 1, then the business-as-normal researcher will outcompete the groundbreaking-insight researcher; as the saying goes, “publish or perish”. Without slack, no researcher can unilaterally escape the system; their best option will always be to continue business as usual.
But group selection makes the situation less hopeless. Universities have long time-horizons and good incentives; they want to get famous for producing excellent research. Universities have rulesets that bind their individual researchers, for example “after a while good researchers get tenure”. And since universities compete with each other, each is incentivized to come up with the ruleset that maximizes long-term researcher productivity. So if tenure really does work better than constant vicious competition, then (absent the usual culprits like resistance-to-change, weird signaling equilibria, politics, etc) we should expect universities to converge on a tenure system in order to produce the best work. In fact, we should expect universities to evolve a really impressive ruleset for optimizing researcher incentives, just as impressive as the clever mechanisms the human body uses to prevent cancer (since this seems a bit optimistic, I assume the usual culprits are not absent).
The same is true for grant-writing; naively you would want some competition to make sure that only the best grant proposals get funded, but too much competition seems to stifle original research, so much so that some funders are throwing out the whole process and selecting grants by lottery, and others are running grants you can apply for in a half-hour and hear back about two days later. If there’s a feedback mechanism – if these different rulesets produce different-quality research, and grant programs that produce higher-quality research are more likely to get funded in the future – then the rulesets for grants will gradually evolve, and the competition for grants will take place in an environment with whatever the right evolutionary parameters for evolving good research are.
I don’t want to say these things will definitely happen – you can read Inadequate Equilibria for an idea of why not. But they might. The evolutionary dynamics which would normally prevent them can be overcome. Two-layer evolutionary systems can produce their own slack, if having slack would be a good idea.
IV.
That was a lot of paragraphs, and a lot of them started with “imagine a hypothetical situation where…”. Let’s look deeper into cases where an understanding of slack can inform how we think about real-world phenomena. Seven examples:
1. Monopolies. Not the kind that survive off overregulation and patents, the kind that survive by being big enough to crush competitors. These are predators that exploit low-slack environments. If Boeing has a monopoly on building passenger planes, and is exploiting that by making shoddy products and overcharging consumers, then that means anyone else who built a giant airplane factory could make better products at a lower price, capture the whole airplane market, and become a zillionaire. Why don’t they? Slack. In terms of those adaptive fitness landscapes, in between your current position (average Joe) and a much better position at the bottom of a deep pit (you own a giant airplane factory and are a zillionaire), there’s a very big hill you have to climb – the part where you build Giant Airplane Factory Part 1, Giant Airplane Factory Part 2, etc. At each point in this hill, you are worse off than somebody who was not building an as-yet-unprofitable giant airplane factory. If you have infinite slack (maybe you are Jeff Bezos, have unlimited money, and will never go bankrupt no matter how much time and cost it takes before you start earning profits) you’re fine. If you have more limited slack, your slack will run out and you’ll be outcompeted before you make it to the greater-fitness deep pit.
Real monopolies are more complicated than this, because Boeing can shape up and cut prices when you’re halfway to building your giant airplane factory, thus removing your incentive. Or they can do actually shady stuff. But none of this would matter if you already had your giant airplane factory fully built and ready to go – at worst, you and Boeing would then be in a fair fight. Everything Boeing does to try to prevent you from building that factory is exploiting your slacklessness and trying to increase the height of that hill you have to climb before the really deep pit.
(Peter Thiel inverts the landscape metaphor and calls the hill a “moat”, but he’s getting at the same concept).
2. Tariffs. Same story. Here’s the way I understand the history of the international auto industry – anyone who knows more can correct me if I’m wrong. Automobiles were invented in the early 20th century. Several Western countries developed homegrown auto industries more or less simultaneously, with the most impressive being Henry Ford’s work on mass production in the US. Post-WWII Japan realized that its own auto industry would never be able to compete with more established Western companies, so it placed high tariffs on foreign cars, giving local companies like Nissan and Toyota a chance to get their act together. These companies, especially Toyota, invented a new form of auto production which was actually much more efficient than the usual American methods, and were eventually able to hold their own. They started exporting cars to the US; although American tariffs put them at a disadvantage, they were so much better than the American cars of the time that consumers preferred them anyway. After decades of losing out, the American companies adopted a more Japanese ethos, and were eventually able to compete on a level playing field again.
This is a story of things gone surprisingly right – Americans and Japanese alike were able to get excellent inexpensive cars. Two things had to happen for it to work. First, Japan had to have high enough tariffs to give their companies some slack – to let them develop their own homegrown methods from scratch without being immediately outcompeted by temporarily-superior American competitors. Second, America had to have low enough tariffs that eventually-superior Japanese companies could outcompete American automakers, and Japan’s fitness-improving innovations could spread.
From the perspective of a Toyota manager, this is analogous to the eyeless alien story. You start with some good-enough standard (blind animals, American car companies). You want to evolve a superior end product (eye-having animals, Toyota). The intermediate steps (an animal with only Eye Part 1, a kind of crappy car company that stumbles over itself trying out new things) are less fit than the good-enough standard. Only when the inferior intermediate steps are protected from competition (through evolutionary randomness, through tariffs) can the superior end product come into existence. But you want to keep enough competition that the superior end product can use its superiority to spread (there is enough evolutionary competition that having eyes reaches fixation, there is enough free trade that Americans preferentially buy Toyota and US car companies have to adopt its policies).
From the perspective of an economic historian, maybe it’s a group selection story. The various stakeholders in the US auto industry – Ford, GM, suppliers, the government, labor, customers – competed with each other in a certain way and struck some compromise. The various stakeholders in the Japanese auto industry did the same. For some reason the American compromise worked worse than the Japanese one – I’ve heard stories about how US companies were more willing to defraud consumers for short-term profit, how US labor unions were more willing to demand concessions even at the cost of company efficiency, how regulators and executives were in bed with each other to the detriment of the product, etc. Every US interest group was acting in its own short-term self-interest, but the Japanese industry-as-a-whole outcompeted the American one and the Americans had to adjust.
3. Monopolies, Part II. Traditionally, monopolies have been among the most successful R&D centers. The most famous example is Xerox; it had a monopoly on photocopiers for a few decades before losing an anti-trust suit in the late 1970s; during that period, its PARC R&D program invented “laser printing, Ethernet, the modern personal computer, graphical user interface (GUI) and desktop paradigm, object-oriented programming, [and] the mouse”. The second most famous example is Bell Labs, which invented “radio astronomy, the transistor, the laser, the photovoltaic cell, the charge-coupled device, information theory, the Unix operating system, and the programming languages B, C, C++, and S” before the government broke up its parent company AT&T. Google seems to be trying something similar, though it’s too soon to judge their outcomes.
These successes make sense. Research and development is a long-term gamble. Devoting more money to R&D decreases your near-term profits, but (hopefully) increases your future profits. Freed from competition, monopolies have limitless slack, and can afford to invest in projects that won’t pay off for ten or twenty years. This is part of Peter Thiel’s defense of monopolies in Zero To One.
An administrator tasked with advancing technology might be tempted to encourage monopolies in order to get more research done. But monopolies can also be stagnant and resistant to change; it’s probably not a coincidence that Xerox wasn’t the first company to bring the personal computer to market, and ended up irrelevant to the computing revolution. Like the eyeless aliens, who will not evolve in conditions of perfect competition or perfect lack of competition, probably all you can do here is strike a balance. Some Communist countries tried the extreme solution – one state-supported monopoly per industry – and it failed the test of group selection. I don’t know enough to have an opinion on whether countries with strong antitrust eventually outcompete those with weaker antitrust or vice versa.
4. Strategy Games. I like the strategy game Civilization, where you play as a group of primitives setting out to found a empire. You build cities and infrastructure, research technologies, and fight wars. Your world is filled with several (usually 2 to 7) other civilizations trying to do the same.
Just like in the real world, civilizations must decide between Guns and Butter. The Civ version of Guns is called the Axe Rush. You immediately devote all your research to discovering how to make really good axes, all your industry to manufacturing those axes, and all your population into wielding those axes. Then you go and hack everyone else to pieces while they’re still futzing about trying to invent pottery or something.
The Civ version of Butter is called Build. You devote all your research, industry, and populace to laying the foundations of a balanced economy and culture. You invent pottery and weaving and stuff like that. Soon you have a thriving trade network and a strong philosophical tradition. Eventually you can field larger and more advanced armies than your neighbors, and leverage the advantage into even more prosperity, or into military conquest.
Consider a very simple scenario: a map of Eurasia with two civilizations, Rome and China.
If both choose Axe Rush, then whoever Axe Rushes better wins.
If both choose Build, then whoever Builds better wins.
What if Rome chooses Axe Rush, and China chooses Build?
Then it depends on their distance! If it’s a very small map and they start very close together, Rome will probably overwhelm the Chinese before Build starts paying off. But if it’s a very big map, by the time Roman Axemen trek all the way to China, China will have Built high walls, discovered longbows and other defensive technologies, and generally become too strong for axes to defeat. Then they can crush the Romans – who are still just axe-wielding primitives – at their leisure.
Consider a more complicated scenario. You have a map of Earth. The Old World contains Rome and China. The New World contains Aztecs. Rome and China are very close to each other. Now what happens?
Rome and China spend the Stone, Bronze, and Iron Ages hacking each other to bits. Aztecs spend those Ages building cities, researching technologies, and building unique Wonders of the World that provide powerful bonuses. In 1492, they discover Galleons and start crossing the ocean. The powerful and advanced Aztec empire crushes the exhausted axe-wielding Romans and Chinese.
This is another story about slack. The Aztecs had it – they were under no competitive pressure to do things that paid off next turn. The Romans and Chinese didn’t – they had to be at the top of their game every single turn, or their neighbor would conquer them. If there was an option that made you 10% weaker next turn in exchange for making you 100% stronger ten turns down the line, the Aztecs could take it without a second thought; the Romans and Chinese would probably have to pass.
Okay, more complicated Civilization scenario. This time there are two Old World civs, Rome and China, and two New World civs, Aztecs and Inca. The map is stretched a little bit so that all four civilizations have the same amount of natural territory. All four players understand the map layout and can communicate with each other. What happens?
Now it’s a group selection problem. A skillful Rome player will private message the China player and explain all of this to her. She’ll remind him that if one hemisphere spends the whole Stone Age fighting, and the other spends it building, the builders will win. She might tell him that she knows the Aztec and Inca players, they’re smart, and they’re going to be discussing the same considerations. So it would benefit both Rome and China to sign a peace treaty dividing the Old World in two, stick to their own side, and Build. If both sides cooperate, they’ll both Build strong empires capable of matching the New World players. If one side cooperates and the other defects, it will easily steamroll over its unprepared opponent and conquer the whole Old World. If both sides defect, they’ll hack each other to death with axes and be easy prey for the New Worlders.
This might be true in Civilization games, but real-world civilizations are more complicated. Orson Welles said:
In Italy, for thirty years under the Borgias, they had warfare, terror, murder and bloodshed, but they produced Michelangelo, Leonardo da Vinci and the Renaissance. In Switzerland, they had brotherly love, they had five hundred years of democracy and peace – and what did that produce? The cuckoo clock.
So maybe a little bit of internal conflict is good, to keep you honest. Too much conflict, and you tear yourselves apart and are easy prey for outsiders. Too little conflict, and you invent the cuckoo clock and nothing else. The continent that conquers the world will have enough pressure that its people want to innovate, and enough slack that they’re able to.
This is total ungrounded amateur historical speculation, but when I hear that I think of the Classical world. We can imagine it as divided into a certain number of “theaters of civilization” – Greece, Mesopotamia, Egypt, Persia, India, Scythia, etc. Each theater had its own rules governing average state size, the rules of engagement between states, how often bigger states conquered smaller states, how often ideas spread between states of the same size, etc. Some of those theaters were intensely competitive: Egypt was a nice straight line, very suited to centralized rule. Others had more slack: it was really hard to take over all of Greece; even the Spartans didn’t manage. Each theater conducted its own “evolution” in its own way – Egypt was ruled by a single Pharaoh without much competition, Scythia was constant warfare of all against all, Greece was isolated city-states that fought each other sometimes but also had enough slack to develop philosophy and science. Each of those systems did their own thing for a while, until finally one of them produced something perfect: 4th century BC Macedonia. Then it went out and conquered everything.
If Welles is right, the point isn’t to find the ruleset that promotes 100% cooperation. It’s to find the ruleset that promotes an evolutionary system that makes your group the strongest. Usually this involves some amount of competition – in order to select for stronger organisms – but also some amount of slack – to let organisms develop complicated strategies that can make them stronger. Despite the earlier description, this isn’t necessarily a slider between 0% competition and 100% competition. It could be much more complicated – maybe alternating high-slack vs. low-slack periods, or many semi-isolated populations with a small chance of interaction each generation, or alternation between periods of isolation and periods of churning.
In a full two-layer evolution, you would let the systems evolve until they reached the best parameters. Here we can’t do that – Greece has however many mountains it has; its success does not cause the rest of the world to grow more mountains. Still, we randomly started with enough different groups that we got to learn something interesting.
(I can’t emphasize enough how ungrounded this historical speculation is. Please don’t try to evolve Alexander the Great in your basement and then get angry at me when it doesn’t work)
5. The Long-Term Stock Exchange. Actually, all stock exchanges are about slack. Imagine you are a brilliant inventor who, given $10 million and ten years, could invent fusion power. But in fact you have $10 and need work tomorrow or you will starve. Given those constraints, maybe you could start, I don’t know, a lemonade stand.
You’re in the same position as the animal trying to evolve an eye – you could create something very high-utility, if only you had enough slack to make it happen. But by default, the inventor working on fusion power starves to death tomorrow (or at least makes less money than his counterpart who ran the lemonade stand), the same way the animal who evolves Eye Part 1 gets outcompeted by other animals who didn’t and dies out.
You need slack. In the evolution example, animals usually stumble across slack randomly. You too might stumble across slack randomly – maybe it so happens that you are independently wealthy, or won the lottery, or something.
More likely, you use the investment system. You ask rich people to give you $10 million for ten years so you can invent fusion; once you do, you’ll make trillions of dollars and share some of it with them.
This is a great system. There’s no evolutionary equivalent. An animal can’t pitch Darwin on its three-step plan to evolve eyes and get free food and mating opportunities to make it happen. Wall Street is a giant multi-trillion dollar time machine funneling future profits back into the past, and that gives people the slack they need to make the future profits happen at all.
But the Long-Term Stock Exchange is especially about slack. They are a new exchange (approved by the SEC last year) which has complicated rules about who can list with them. Investors will get extra clout by agreeing to hold stocks for a long time; executives will get incentivized to do well in the far future instead of at the next quarterly earnings report. It’s making a deliberate choice to give companies more slack than the regular system and see what they do with it. I don’t know enough about investing to have an opinion, except that I appreciate the experiment. Presumably its companies will do better/worse than companies on the regular stock exchange, that will cause companies to flock toward/away from it, and we’ll learn that its new ruleset is better/worse at evolving good companies through competition than the regular stock exchange’s ruleset.
6. That Time Ayn Rand Destroyed Sears. Or at least that’s how Michael Rozworski and Leigh Phillips describe Eddie Lampert’s corporate reorganization in How Ayn Rand Destroyed Sears, which I recommend. Lampert was a Sears CEO who figured – since free-market competitive economies outcompete top-down economies, shouldn’t free-market competitive companies outcompete top-down companies? He reorganized Sears as a set of competing departments that traded with each other on normal free-market principles; if the Product Department wanted its products marketed, it would have to pay the Marketing Department. This worked really badly, and was one of the main contributors to Sears’ implosion.
I don’t have a great understanding of exactly why Lampert’s Sears lost to other companies even though capitalist economies beat socialist ones; Rozworski and Phillips’ People’s Republic Of Wal-Mart, which looks into this question, is somewhere on my reading list. But even without complete understanding, we can use group selection to evolve the right parameters. Imagine an economy with several businesses. One is a straw-man communist collective, where every worker gets paid the same regardless of output and there are no promotions (0% competition, 100% cooperation). Another is Lampert’s Sears (100% competition, 0% cooperation). Others are normal businesses, where employees mostly work together for the good of the company but also compete for promotions (X% competition, Y% cooperation). Presumably the normal business outcompetes both Lampert and the commies, and we sigh with relief and continue having normal businesses. And if some of the normal businesses outcompete others, we’ve learned something about the best values of X and Y.
7. Ideas. These are in constant evolutionary competition – this is the insight behind memetics. The memetic equivalent of slack is inferential range, aka “willingness to entertain and explore ideas before deciding that they are wrong”.
Inferential distance is the number of steps it takes to make someone understand and accept a certain idea. Sometimes inferential distances can be very far apart. Imagine trying to convince a 12th century monk that there was no historical Exodus from Egypt. You’re in the middle of going over archaeological evidence when he objects that the Bible says there was. You respond that the Bible is false and there’s no God. He says that doesn’t make sense, how would life have originated? You say it evolved from single-celled organisms. He asks how evolution, which seems to be a change in animals’ accidents, could ever affect their essences and change them into an entirely new species. You say that the whole scholastic worldview is wrong, there’s no such thing as accidents and essences, it’s just atoms and empty space. He asks how you ground morality if not in a striving to approximate the ideal embodied by your essence, you say…well, it doesn’t matter what you say, because you were trying to convince him that some very specific people didn’t leave Egypt one time, and now you’ve got to ground morality.
Another way of thinking about this is that there are two self-consistent equilibria. There’s your equilibrium, (no Exodus, atheism, evolution, atomism, moral nonrealism), and the monk’s equilibrium (yes Exodus, theism, creationism, scholasticism, teleology), and before you can make the monk budge on any of those points, you have to convince him of all of them.
So the question becomes – how much patience does this monk have? If you tell him there’s no God, does he say “I look forward to the several years of careful study of your scientific and philosophical theories that it will take for that statement not to seem obviously wrong and contradicted by every other feature of the world”? Or does he say “KILL THE UNBELIEVER”? This is inferential range.
Aristotle supposedly said that the mark of an educated man is to be able to entertain an idea without accepting it. Inferential range explains why. The monk certainly shouldn’t immediately accept your claim, when he has countless pieces of evidence for the existence of God, from the spectacular faith healings he has witnessed (“look, there’s this thing called psychosomatic illness, and it’s really susceptible to this other thing called the placebo effect…”) to Constantine’s victory at the Mulvian Bridge despite being heavily outnumbered (“look, I’m not a classical scholar, but some people are just really good generals and get lucky, and sometimes it happens the day after they have weird dreams, I think there’s enough good evidence the other way that this is not the sort of thing you should center your worldview around”). But if he’s willing to entertain your claim long enough to hear your arguments one by one, eventually he can reach the same self-consistent equilibrium you’re at and judge for himself.
Nowadays we don’t burn people at the stake. But we do make fun of them, or flame them, or block them, or wander off, or otherwise not listen with an open mind to ideas that strike us at first as stupid. This is another case where we have to balance competition vs. slack. With perfect competition, the monk instantly rejects our “no Exodus” idea as less true (less memetically fit) than its competitors, and it has no chance to grow on him. With zero competition, the monk doesn’t believe anything at all, or spends hours patiently listening to someone explain their world-is-flat theory. Good epistemics require a balance between being willing to choose better ideas over worse ones, and open-mindedly hearing the worse ones out in case they grow on you.
(Thomas Kuhn points out that early versions of the heliocentric model were much worse than the geocentric model, that astronomers only kept working on them out of a sort of weird curiosity, and that it took decades before they could clearly hold their own against geocentrism in a debate).
Different people strike a different balance in this space, and those different people succeed or fail based on their own epistemic ruleset. Someone who’s completely closed-minded and dogmatic probably won’t succeed in business, or science, or the military, or any other career (except maybe politics). But someone who’s so pathologically open-minded that they listen to everything and refuse to prioritize what is or isn’t worth their time will also fail. We take notice of who succeeds or fails and change our behavior accordingly.
Maybe there’s even a third layer of selection; maybe different communities are more or less willing to tolerate open-minded vs. close-minded people. The Slate Star Codex community has really different epistemic norms from the Catholic Church or Infowars listeners; these are evolutionary parameters that determine which ideas are more memetically fit. If our epistemics make us more likely to converge on useful (not necessarily true!) ideas, we will succeed and our epistemic norms will catch on. Francis Bacon was just some guy with really good epistemic norms, and now everybody who wants to be taken seriously has to use his norms instead of whatever they were doing before. Come up with the right evolutionary parameters, and that could be you!
Last part of “4. Strategy Games” seems related to the end of Guns, Germs, Steel (see last paragraph of Synopsis -> Outline of Theory). https://en.m.wikipedia.org/wiki/Guns,_Germs,_and_Steel I think Jared Diamond’s argument there is that less slack due to geography-induced competition led to critical bad policy decisions being limited in duration, among other things
Yeah, I thought of that too.
https://en.wikipedia.org/wiki/Guns,_Germs,_and_Steel#Outline_of_theory
Slack: Getting Past Burnout, Busywork, and the Myth of Total Efficiency
As I recall, this is more about how much competition a company is expecting rather than the actual level of competion, but it’s about the fear of competition getting so extreme that people in the company don’t have time to think about their decisions or provide quality products.
The first example is a bit unsatisfying, since an actual eye is far too complicated to evolve by chance, even if you have all the evolutionary slack in the world. Still, the insight that slack lets you get out of a local minima is a good one. In physics, temperature plays the same role, allowing systems escape to lower energy minima.
An important ingredient many of the other examples seems to be good intentions. We can imagine a completely selfish scientific researcher. With no slack, they will be forced to churn out trendy paper after trendy paper in order to keep their jobs. If they have a lot of slack, perhaps bulletproof tenure with no chance of being fired, they will spend their time playing video-games. If the amount of slack is intermediate, they will churn out enough trendy papers to keep their jobs, and spend the remaining time playing video-games. The only time that groundbreaking research is a possibility is when we have a scientist who is intrinsically motivated to produce good science. That is when adding more slack starts to produce better outcomes. Similarly, while an enlightened monopolist might choose to invest in R&D, other monopolists will instead choose to pay themselves a hefty bonus and buy a yacht. The correct amount to slack to have depends on how enlightened your monopolists tend to be. It seems like slack is playing a very different role here than in the evolutionary case, where it is channeled into pure randomness.
Which kind of eye ? I was under the impression that the evolution of vision was fairly well understood, but there are many different kinds of eyes out there, so I could be wrong. But, as Scott says, the actual eye is still a poor example, because each step of its evolution does confer an advantage — unlike the steps required to build the (hypothetical) giant airplane factory. So, real eyes require very little slack compared to airplane factories.
I assume the three “eye parts” are more something like “photoreceptive chemical” and “two expression sites on two sides of the organism” and “differential behavior to turn towards the light”. I’d guess IRL this was bypassed by having a preexisting turn-towards-chemical-that-indicates-food behavior. Maybe they use a different food-seeking mechanism that can’t be reused like that.
But of course the actual eye did evolve by chance, if we mean a series of random mutations that eventually led to what we’ve got. Our eyes are different than the eyes of a housefly or a squid, and no entity intentionally designed any of them.
Scott’s example is satisfying, because he’s asking us to imagine that instead of the intermediate steps conferring a small benefit (as they did on Earth, which is why they stuck around long enough to step by step become what we’ve got), they would only be a burden.
The notion that evolution works by pure chance is a common misconception. Selection for various traits is a vitally important ingredient. Take that away, as Scott does, and you won’t end up with an eye. Maybe if evolving an eye only required 3 single-letter mutations, you could get there by chance, even if the first two were deleterious, but in real life the required number of mutations is much larger.
Wait, what distinction are you making here? Scott doesn’t ignore selection — hence “Here on Earth, scientists believe each of these mutations must have had its own benefits”.
[edit]
I think I get your point: eyes are too complex to evolve without incremental fitness benefits; if you assume they are only useful when complete, then the odds are prohibitively against evolution ever discovering them, slack notwithstanding. (The ‘common misconception’ bit threw me, as I don’t think Seth or Scott are making that mistake.)
The eye example is just a simplified metaphor, though. To make it more realistic, you could increase the resolution: let’s say that milestones 1-1000 on the road to an eye are individually fitness-increasing, but each milestone requires multiple (though not an impossibly large number of) mutations that are individually useless. Now slack has a plausible role to play.
sethgodin makes it sound like the road to an eye is just as likely as any other road. While chance played a role, it was not the sole driver.
There are real-world examples of such hillocks which evolution cannot overcome. Staying with the eye example: Evolution will never fix our retina which is sitting the wrong way: The nerve ends don’t come out at the back, towards the brain, but in the front, and we need the blind spot of Mariotte to let them pass through. Squids don’t have this problem (which by the way indicates that their eye seems to have evolved independently). Evolution is unable to turn the whole retina around!
Another example: Evolution has never invented the wheel. There are tumbleweeds or some rolling bugs and caterpillars, but no wheels turning around an axis. This is a concept which is hard to invent by evolution. Of course the concept of irreducible complexity is unscientific, because it is an argument ex post: If there were wheeled animals, we wouldn’t assume that they are impossible.
The problem with this example is that wheels are actually not very adaptive. Even in today’s world with many pre-existing man-made roads there aren’t many organisms you would purposefully design with wheels instead of legs. Think about how annoying it is to get around in a wheelchair and now imagine you need the chair to catch prey, flee from predators, or navigate a plant-rich environment.
It’s also not at all clear that evolution would be incapable of producing a wheel if there were appropriate selection pressure. The Flagellum is remarkably similar to an actual rotor and that has been produced by evolution.
This is such absolute and total horseshit you should be ashamed to have even thought it, much less posted it on a public forum.
The evolution of vision is extremely well documented, numerous cases of EXTANT intermediate forms exist, the developmental biology is far simpler than you know (which is obvious, since you clearly know zero biology), and every step does, in fact, produce an advantage.
Photoreceptive pigments are extremely common throughout life, even in bacteria, mediating biased random walk patterns to avoid or stay in light. Distributed photoreceptors give gradients, making motion more efficient. Concentrated patches allow extremely rudimentary image formation (though some species with whole-body photoreception can still form images, like some echinoderms). Making the patch concave improves focus, until you eventually get a pinhole camera (a widespread form in many worms, and convergently evolved in the “second eyes” of vipers, i.e. their heat pits). A supporting body of fluid in the core can, with easy modification, become a lens, and bingo, the eye, simple and easy. The fact that eyes seem to have evolved multiple times independently shortly after the origin of multicellular life is proof enough that, far from impossible, it’s quite easy to evolve an eye – it’s evolved more times than powered flight, after all.
I’m not positive, but I’m pretty sure that phi is not disputing this.
This is the key point. I believe in phi’s statement that “an actual eye is far too complicated to evolve by chance,” the emphasis is not on “evolve”, but on “by chance.” An eye could evolve by natural selection, just not by chance alone, as in Scott’s example. Because the eye has many intermediate stages, not just three, it couldn’t evolve without each of those stages conferring some advantage of their own; getting all of them at once just by chance would be unlikely enough that it would never happen. Because each state does, in the real world, confer some advantage, the eye can, of course, evolve in the real world; it’s just that it makes Scott’s example a bit forced, because in his example he postulates that the stages don’t confer any advantage until you have all of them. At least, that’s how I read phi’s comment.
“Evolution” does NOT mean “spontaneous occurrence of mutations”. It means “change in allele frequency in a population over time”. Either your interpretation is wrong and phi’s statement is as stupid as I think, or phi mis-used the word so badly it’s like calling a train a boat.
Less of this, please.
I think a careful reading and ten seconds of thought about phi’s comment will reveal that they aren’t actually trying to argue that IRL eyes were created by divine intervention, as you seem to thing they’re saying. The point phi’s trying to make is that, if each individual part of an eye did not confer an advantage (i.e. in Scott’s hypothetical alien species), then complex eyes would not evolve even on evolutionary timescales due to their improbability.
And even if they were advocating creationism, saying “you should be ashamed to have even thought it” is…at odds with the culture we try to foster here. “I’m astounded you actually believe that, could you explain why?” tends to lead to more interesting discussions.
I don’t think you need any mystical concept of “intrinsically motivated to produce good science”. All you need is the idea that a certain number of trendy papers can keep your job, while a slightly larger number gets you a bit of a promotion, and an individual revolutionary paper gets you nothing, while a revolutionary paper that gets followed up by a bunch of trendy papers on its trend gets you a named research chair at Caltech. If you tune the “you’re fired” parameter too close to the margins, then everyone does just trendy papers. If you tune it too low, then everyone does just a few trendy papers and then spends all the rest of the day attempting their solitary revolutionary things, and no one gets the boost of other people’s trendy papers. If you tune it just right though, then everyone attempts a few revolutionary papers and then does enough trendy papers that some people are able to strike the jackpot.
Companies don’t evolve, though. If they did, they probably wouldn’t be so commonly dysfunctional as they are…
I don’t find this a very interesting disagreement. Corporations clearly do something evolution-like, but it clearly isn’t exactly the same as biological evolution. Whether or not you call that “evolution” is a stylistic choice, and I think mine is appropriate here, especially since everyone now seems to agree that saying “cultural evolution” is okay.
This is a great post, Scott, worth the effort you put into it. Twenty years ago, I spent a year of my life writing an entire book on the topic, which is my least successful: https://www.amazon.com/Survival-Not-Enough-Shift-Happens/dp/0743221230
People have a hard time understanding the massive time scale associated with biological evolution and their eyes glaze over (I’ve seen it from the stage, almost every time) when it’s talked about.
I think that memetics, red queens, sexual selection and the rest are fascinating and accurate metaphors for so much of what builds our culture. Thanks for highlighting it here.
I don’t think that’s so clear at all. As Eliezer points out, the entire descent part is missing. Companies’ descendants don’t really manage to replicate their parents, because so much of what makes a company tick is really hard to transmit! So, it isn’t transmitted. And, of course, “more successful” for a company means more money, not more descendants.
On the other side of things, of course, companies do, like, try to optimize themselves, which is also quite un-evolution-like.
I think the resemblance to evolution is wholly superficial. It just really isn’t evolution. There’s some optimization going on, and some competition, but those are basically the only two aspects that match. Evolution’s whole mechanism for gradual improvement is gone.
And like — I think this is why large companies are still so crap. If corporate evolution were a real thing, I don’t think this would be such a consistent problem. But instead we have to rely on the more legible sorts of optimization, which normally I like but in this case don’t seem to do very well, because people don’t really know how to write down or transmit a corporate culture, including mechanisms to prevent a company from falling into mazedom that won’t themselves be perverted by the suits to drive the system to mazedom more quickly.
Cultural evolution obviously isn’t the same as biological evolution either, but I think of “cultural evolution” as more of just a phrase describing how cultures change rather than by a claim that the mechanism is actually similar. Cultures do seem better at replicating themselves into their descendant cultures than companies do (because cultures don’t discretely fission like that). But cultural evolution as a parallel to biological evolution? Yeah, that doesn’t happen. Once again, you have competition, and this time you have something more like descent with modification, but the other aspects are still missing (again, more success doesn’t mean more descendants).
But it isn’t. New corporations are not generally formed by fresh-out college graduates (or dropouts), but by people who have worked for existing corporations. They take the best practices of the firms they have worked for and duplicate them in the new companies they found.
And for that matter, existing companies often attempt to duplicate the best practices of their own competitors, or of firms in related but not directly competing industries, when they feel that their own practices are lacking in some respect. So you get Darwinian evolution, Lamarckian evolution, and Intelligent Design all working in parallel.
Sounds more like a memetic evolution of business practices, rather than of companies.
@ignamv
A somewhat analogous claim would be saying that organisms don’t evolve, genes do.
Corporations do horizontal gene transfer.
But the selection pressure is not measured simply in profits. Sure, that’s important, but humans are very picky and diverse group, and they can keep many kinds of corporations alive. From Amazon to Patagonia, from the sweatshops of Bangadlesh and the semi-robotized paramilitary Foxconn campuses to the super-exclusive Koenigsegg automotive of Sweden.
But even if we just look at Samsung, Apple and Google, they all have phone brands, they all have phone serious phone software and hardware development departments, they all have a lot of other businesses that make them money, yet basically all are completely different. (Samsung is missing from FAANG, because it’s not US based, Apple is not though of as an Internet business but more as a device maker, Samsung is not a pure tech business, because it’s more like Korea’s engineering department that fall in love with rounded corners.)
Just as human minds are incredibly diverse corporations are too, that doesn’t mean there is no evolution, just means that the causal pathways are long and chaotic.
(Reproduction, heredity, variation in fitness, variation of traits are the required phenomena. We obviously have variation in both categories, and we can say that we have heredity too, but this doesn’t require the corporation to die and reproduce, it’s enough if it changes, incorporates new ideas, or the board replaces the CEO, or the board approves a new approach to doing business; and we have reproduction too, successful corporations grow and copycats emerge inevitably, and they even usually enter into new market segments, acquire competitors, merge with them, and now we have more of the same. That’s reproduction even if it does business as one big entity.)
I regret not expanding upon it at the time, but this is the exact point (and link, even) that I was alluding to at the end of my comment on BR:SooS.
The difference is critical, maybe enough to undermine the entire thesis. As the engineer’s saying goes, a difference in magnitude is a difference in kind – these different selectors may all be called “evolution”, but I’m counting five or six orders of magnitude between the timescales at a minimum.
DownUpthread John comments that “you get Darwinian evolution, Lamarckian evolution, and Intelligent Design all working in parallel” and I agree, but if you’re intuiting how selection pressures produce adaptation then those three do not belong in the same bucket.I really don’t get what that article is on about. Even if you take the worst case assumption that corporate culture is utterly inflexible after creation (which is really isn’t, companies reinvent themselves all the time) there’s still room for evolution in the corporate world. No one is going to copy Juicero’s business model but everyone is copying Uber’s. That is evolution and heritability, even if the genes in question are business memes.
What? The article isn’t talking about how corporate cultures don’t change. It’s talking about how corporate cultures don’t replicate themselves. There’s change, sure, but the optimization process of the successful ones replicating themselves doesn’t happen; at most, other companies try to imitate the successful ones — but the problem is, so much of what makes a company successful is hard to transmit.
I mean… humans don’t replicate themselves and we still evolve. I’m not really following this analogy. Sure corporations don’t obey strict Mendelian principals but neither does 99% of the natural world.
The eye evolution is actually an equivalent of simulated annealing.
Yeah, I kept expecting that to come up given that Scott lives in the very-ML-literate Bay Area.
Yeah, also expected that to be mentioned.
Or, indeed, actual annealing.
Feels like another great place to insert the complete stance. We can slice the cake and come up with many 2 dimensional slices that imply a tradeoff between several competing tensions, some cleaner than others. One question is, why isn’t this stance obvious? Why do thousands of pages need to be written about it with only some people getting it even then? When we ask what is the ideal balance point and get back an answer of ‘it depends’ why does it bother us so much? Because the payoff for finding an actual invariant is so damn high. If you do manage to find metastability you get compression and modularity, which then lets you abstract away that particular tradeoff and build much more complicated things. Things like commodity futures markets that prevent drastic swings of over and undersupply of agriculture in your civilization because farmers now get pricing signals. Things like being able to manufacture complex goods with long supply chains. You have probably heard that premature optimization is the root of all evil. Abstraction is a form of optimization. Premature abstraction seems to be how we make progress, despite there being a hundred false alarms for every fire, because if it’s a false alarm you just go back to work, but if it’s a fire it spreads. Occasionally we abstract away human suffering and get monstrous results. Increasingly suggestive hints from biology that cancer happens not because cells get ‘old’ but because intercellular signaling breaks down, undermining the cooperate-cooperate equilibrium. Making our suffering legible to one another can also get coopeted by moloch, but the payoff if we succeed in extracting invariants would be so damn high.
Slight correction: the cuckoo clock speech wasn’t written by Greene (though it really isn’t out of voice for him at all), but rather added by Orson Welles during shooting when they found they needed another sentence for timing. One of the all-time great films.
https://en.wikipedia.org/wiki/The_Third_Man
And of course, the longer speech it’s part of is one of the most evil speeches in all fiction.
– Have you ever seen any of your victims?
– Do you know, I don’t ever feel comfortable on these sort of things… Victims?
(He opens the door of the Ferris wheel carriage.)
Don’t be melodramatic. Look down there… would you feel any pity if one of those dots stopped moving forever? If I offered you £20,000 for every dot that stopped – would you really, old man, tell me to keep my money? Or would you calculate how many dots you could afford to spare? Free of income tax, old man, free of income tax. It’s the only way to save money nowadays.
I think that it’s a bit inaccurate to refer to dialogue that someone writes for one of their characters as something they “said”. Just because you have one of your characters say something, doesn’t mean you’re endorsing it.
About Switzerland producing only a Cuckoo clock… Switzerland has the most Nobel prizes per capita among decent-sized countries. Einstein lived and worked in Switzerland when he wrote his most famous works.
Also Cuckoo clocks were invented in Bavaria.
The Swiss also used to be Europe’s most in-demand soldiers during the early modern period. And they didn’t have only brotherly love for each other.
Cuckoo clocks stereotypically are produced in Schwarzwald. Which sounds sort of similar to Switzerland, and that’s probably the cause of this mistake.
You can’t look at modern state boundaries when you’re talking about these things. They’re more a reflection of Napoleonic and Bismarkian machinations than actual culture or history.
Augsburg, germanophone Swizerland and the Schwarzwald were all part of the stem dutchy of Schwabia, centred on the Bodensee, under the early HRE. This area was mostly inhabited by the Alemanni/Suebi, who broke through the Limes in late antiquity.
So Augsburgers are culturally (and linguistically) closer to the Swiss Germans, Baden-Württembergers and Alsatian Germans than Bavarians.
Even today, Schwabia is an important region for engineering. It’s the home of Mercedes-Benz/Daimler, Porsche and Bosch.
Also, five hundred years of democracy, peace and brotherly love has got to be worth something by itself.
Prior to salt iodization Switzerland had a cretinism problem.
This concept is why I am usually always in favor of people trying new things and experimenting – no matter how good I think the actual idea is. Uniformity scares me more than anything.
Freeman Dyson has this really interesting bit in one of his books where he says that research of space programs and nuclear programs stalled completely because they were solely government controlled and the government cannot afford to take risks and lose face on “big important things” – it has no slack. The private sector is much better suited for research in a novel field where lots of things need to be tried before we have a good idea of what works.
That’s interesting, because I would have assumed the government has more slack since it can just get into debt endlessly and doesn’t need to be profitable. The government can “lose face” for decades on end but keep doing the same thing anyway. On the other hand, corporations that get bailed out by the government don’t necessarily have to be profitable either. There are a lot of just-so type arguments we could make here.
I think slack creates the opportunity for big important things and big stupid things, everything from impressive marvels of infrastructure, to complete boondoggles that cost billions upon billions but fail to work properly. Slack creates the opportunity for smart people to not have to worry about the shorter time horizon of surviving, but simultaneously you have to wonder about the competence of someone who hasn’t ever faced constraints. Maybe this means that the government should be the slack component and use the endless money bag for long timeframe projects, but it should hire those tested by the constraints of profitability.
On the other other other hand, I do recall something posted here on SSC about how mixed public-private partnerships are actually bad since they corrupt motives. It concerns me that on this topic we could just other hand all day, and there’s no natural isolated experiment to appeal to. All the evidence says we need some compromise between the two factors, but there’s absolutely no telling what that compromise looks like. It seems kind of hopeless.
The government has less financial pressure, but requires more acceptance. Politicians want to be re-elected, just like CEOs want to keep their jobs, but the requirements to keep their job are different.
Even financially, the pressures are different, because people get upset over how the government is spending our money, while people care way less about how companies are spending their money. So the government loses face when they mess up an IT project and waste money, but companies can mess up whatever they want as long as their prices stay reasonable, although they do lose face when they mess around with our data.
So I would say that you are both right, the government both has more and less slack, depending on what slack you look at. Slack is not something you have more or less of, but is is something that you have more or less of in a certain dimension.
You can compare it to a Jewish person having more slack to make Jew-deprecating jokes and a black person having more slack to make black-deprecating jokes, but the Jewish person better not make those same black-deprecating jokes.
To use this theory as a guide to behavior is going to either require plotting out this multi-dimensional slack matrix (in which case: what if the Office of Slack becomes too slack?), or accepting that it subjectively depends on the lens you use what even counts as slack to begin with, making the whole concept less useful.
On the “acceptability” vs “costs” metric. This means its often much easier for a government to plow more and more money into a program that’s already approved/popular, than to do a new program
What do you mean by “the government” in this context? Because short of a major war, nothing is ever done by “the government” as a whole. And even in a major war, no decision is made by “the government” as a whole.
The administration, has to avoid losing face in the eyes of its supporters at least, or they’ll be voted out of office at the next election. And the bureaucrats making operational decisions, if they lose face, will see their career paths hit a brick wall and their departments’ budget and scope diminished at least in relative terms.
But it’s a nestled thing where slack on one level provides slack for another. You go on to say that the civil servants can see budget cuts, but the degree to which this happens is dictated by the administration. It’s not particularly important whether the administration directly runs everything, which would be silly.
That depends on if voters are able to not vote for that party or even that incumbent. In the American system, most disillusioned voters still vote for their party. Liberals pull the lever for Democrats, and conservatives pull the lever for Republicans, while they both try to motivate some portion of the independents. This means there’s a very high bar for policies and leaders becoming untenable. Sure if a Republican or Democratic President is bad enough, the voters from the other party might mobilize slightly more than they otherwise would, but this is happening at the margins, since voters from the same party will largely keep voting for the terrible President because of “the lesser of two evils” logic. This creates a high bar for mass rejection of a leader, and then add in incumbency advantage, and you have a large amount of slack.
In theory, but Republicans rarely slash and burn as much as they say they’re going to. They increase the national debt the same as Democrats do (and before the stimulus under Obama, their record was worse). That’s not to say that things won’t get cut back in one area while the DoD is given even more money, but overall things don’t get obliterated by cuts that often, and this is when the party that supports a competition based worldview is in power.
What you’re talking about are factors that relatively speaking wobble the amount of slack up and down a bit, but the government of the USA and most Western countries are operating with an incredible total amount of slack, in which finances practically don’t matter at all, because mostly voting somebody out doesn’t constrain the government to a scarcity respecting regime. There is a large degree of continuity between parties.
Obviously there’s a limit, but I think it’s a lot higher than you are giving credit for, and this better explains why so many modern governments have in fact been incredibly loose with deficit spending. You could say that they chose this option instead of higher taxes, which is an example of a pressure changing their behavior, but it just means that they found a get out clause. It doesn’t matter if a lion was initially hunting gazelle if its behavior over the last decade has involved having food thrown at its face by keepers. You had competition and it relaxed completely, so now you have slack.
Possibly I’m missing something, but it looks like your entire response is about the willingness of governments (in whole or in part) to spend lots of money they don’t have. Which, yes, they are. But that’s not the problem I am talking about, and I don’t think it is what Scott etc are talking about.
The problem is that “the government” is unwilling to take risks , or at least risks whose downside involves significant loss of face. That’s where there isn’t enough slack, and government action is constrained to doing things that have proven safe in the past. Spending ten billion dolllars on something where we’re guaranteed to muddle through is OK, spending one billion on something that will either succeed or fail spectacularly, not OK.
Given five billion-dollar proposals that each have a 50-50 chance of spectacular success or spectacular failure, where the successess will exceed the ten-gigabuck muddle, and we’ll spend the ten gigabucks. The other plan has a 97% chance of giving better results at half the cost, but it requires five people to wager their prestigious and lucrative careers on a coin toss. You’ll get people who say they are willing to do that, but really they’re each going to hit you up for ten billion and then muddle safely through.
From my experience of talking to people who work in government, there’s a pretty strong selection pressure not financially but for “justifiability”. For a new program to be funded it needs to be the sort of thing that you can convince people (depending where in the system you are that might mean management, politicians, interest groups or the public writ large) that it’s the sort of thing that you should do. And it’s always much easier to justify the status quo than any change, however good it is in theory.
Education is a good example of this. There’s lots of research on ways to do it better, (I’m going to avoid mentioning any specifics so thus doesn’t get derailed, but insert your favorite here) but any of them would require a major change to the status quo of “what schools look like” and its difficult to convince the relevant stakeholders (eg parents, government) that they are worth doing, even if the experts agree they are, think of the massive backlash to any change made in the American education system in the last few decades. Education is particularly bad because it has a large stakeholder group (parents and the public in general) who aren’t esperts, but think they understand the system and how it should work due to having spent 10+ years in it. And the outcomes are long term and hard to measure. (If you institute “magic new curriculum x” even if its better on every metric, you won’t see the results for years or decades). And people have strong feelings about anything that affects their children. So large scale change is very difficult. Only happening with either broad public support or a big top down effort from government willing to sacrifice short term political benefit for the long term outcome.
Your example is rather poor, since the ‘experts’ on education reform seem to typically be extremist ideologues, with extremely little evidence that their proposed reforms are better (let alone much better), yet with very high confidence that they are.
This is not true (e.g. https://eric.ed.gov/?id=ED397535 ). Talking heads aren’t experts – everyone knows this at some level.
It’s a bit politically incorrect, but wealth inequality is a major source of slack. A private company that’s held by millions of investors who are all relying on the stock for their retirement has little slack. A company wholly owned by an eccentric billionaire has lots.
Good point/
But it isn’t necessary that the eccentric billionaire own the company outright, or even that he have a majority stake, So long as it’s clear that it would take an implausibly coordinated shareholder revolt to divert the company from the eccentric billionaire’s chosen path, the rest of the investors will be people who are happy with eccentric-billionaire management.
iPhones exist because Steve Jobs was a billionaire. With just pension funds and their conservative CEO picks, you’d all be making due with Blackberries.
Have space programs really stalled? Seems that we keep doing the important things (GPS, meterology, research satellites) and we just abandoned the flag-waving part (sending humans to distant celestial bodies).
We were doing all those things forty years ago. We’re doing them somewhat better now, but only by small marginal improvements. What we are conspicuously not doing, is anything whose failure would represent a conspicuous loss of face for the people in charge. GPS satellites are useful, but if one of them fails you’ll probably never notice it. If an entire block of satellites turn out to be lemons, you’ll probably never notice.
See also the James Webb Space Telescope. Originally proposed as a quick $500E6 project because with fancy modern tech we can way outdo Hubble at bargain prices – but it would mean taking “unnecessary” technical risks with a high-profile flagship program. So, enter plan B where we spend ten billion dollars and run the schedule out an extra fifteen-plus years until all the people who set it up are safely retired.
Also, count the number of “return to the Moon” proposals that have had their budget and schedule grown to the point that they are cancelled before anyone has to take the risk of launching an astronaut on a new spaceship.
The cost per kg to orbit, and the cost and time to build a new launch vehicle, did not improve much in quite a few decades (and has sometimes gone backwards, e.g. the Space Launch System). Several companies in the U.S. tried to improve things and all failed, some asserting that government policy made it impossible to succeed.
SpaceX shows how much room there was to improve. First they got to orbit on a shoestring. Then they became competitive in just a few years. Then they developed reusable orbital-class boosters and dominated the market. Now they’re building a fully-reusable Apollo-class vehicle for $2-3 billion and a few years of R&D in tents, and projecting 100 tons to orbit for $2M per flight.
So yes, we have maintained a space capability, but it seems a case study in stagnation until SpaceX came along.
I agree whole heartedly with your comment, but when one says “SpaceX came along” that really means “Elon Musk risked half his large fortune against everyone’s advice and worked really hard with lots of brilliant people for 15 years” (the other half went toward Tesla). The human race could have very easily seen no advancement in space launch capabilities for the foreseeable future. Maybe China?
Elon Musk’s companies have never been first to market. This is probably very important, in that they didn’t have sunk costs on ideas that failed previously (or merely on old tech that worked), and could see market ‘pain points’ (I hate that term) in advance, and directly address those points.
https://en.wikipedia.org/wiki/Space_launch_market_competition
Did Scott just coin the term “inferential range”? I think that’s a keeper.
If you tell him there’s no God, does he say “I look forward to the several years of careful study of your scientific and philosophical theories that it will take for that statement not to seem obviously wrong and contradicted by every other feature of the world”? Or does he say “KILL THE UNBELIEVER”?
He’s probably likely to say “Who do you think you are, Peter Abelard?” and then refer you to the Cluniacs 🙂
“Inferential distance” has been a term used on Less Wrong for a while
https://wiki.lesswrong.com/wiki/Inferential_distance
There’s an important limitation to the fitness-landscape model – it assumes that evolution-as-a-dynamical-system has a potential function. In other words, it assumes that any given point in state space has an innate “stability”, with less-stable states yielding to more-stable ones. But this need not be the case (and in mathematical terms, it would be kind of surprising if it did).
Consider a system with two (self-interested) agents, A and B. A and B have come into possession of a cursed crown. While wearing it, one feels intense agony, but can compel others to put on the crown instead. Once the crown is off of you, the agony goes away, but so does the power to compel others.
The crown begins on A’s head. A does not want to be in pain, so A uses the crown’s power to compel B to put it on instead. Now A’s pain is relieved. B is now in agony, so B uses the crown’s power to put it on A’s head. And so on and so forth for eternity.
Actually, this is just a simplified version of foxes and rabbits. In a world with no mutations, Moloch (here interpreted as simply the god of dynamical systems) leads the foxes-and-rabbits system around in circles for eternity: high foxes, low rabbits -> low foxes, low rabbits -> low foxes, high rabbits -> high foxes, high rabbits -> repeat. This system happens to have an equilibrium (a population of foxes and rabbits that perfectly balance forces out), but some systems (like the cursed-crown one) don’t. Some systems need not even have stable cycles.
The implication of this is that unlike a true fitness landscape, the amount of work to get from point A to point B in such a system’s state space is (potentially unboundedly) path-dependent. And I don’t just mean in the sense that the “humps” you have to cross are path dependent, I mean in the sense that the height-difference is in some sense path dependent. Like an Escher drawing, there can be a “downslope” from A to B and from B to C *and from C back to A*, while the A -> C -> B -> A route is entirely uphill, something that is impossible in a true fitness-landscape model.
Relatedly covering lack of stable cycle, Nisan’s excellent book report on Theory of Games and Economic Behavior.
It is true that potential-based views can be misleading, but even if attractors are not fixed points, they can still have basins of attraction. Will you end up in the A-C-B cycle (or strange attractor or whatever), or the D-E-F, or somewhere else, and can you jump from one to another?
So it’s not so clear-cut, and you can use ideas like Morse decomposition to separate a dynamics into “potential-like behavior far from the attractors” (falling into a basin), and “non-potential-like around the attractors” (A-C-B-A)
What you’re talking about is well-understood in evolutionary biology, particularly in the context of frequency-dependent selection, in which the rare phenotype is always favored. These type of systems tend to be either a) sexual preference, b) exploiting fluctuating food types, or c) immunological arms races with parasites or other pathogens.
In contrast, fixed fitness landscapes (or mostly fixed) tend to occur with more mechanistic systems, such as locomotion, feeding, armor, etc., where the underlying landscape is mediated not by other biological factors, but by physics.
“particularly in the context of frequency-dependent selection, in which the rare phenotype is always favored”
I think you mean “particularly in the context of frequency-dependent selection in which the rare phenotype is always favored”
I’m skeptical about your analogy in Part 1.
In the early days of machine learning, people worried a lot about local minima. You optimize 10 parameters one step at a time, you might not reach the absolute best configuration (the ‘global minimum’), same as the picture with the ‘puddle’ and ‘pit’ you showed.
But nowadays ML has stopped worrying about local minima. Why? Because we’re not fitting 10 parameter models, we’re fitting 10 million parameter models. And the 2D or 3D intuitions don’t quite carry over.
Imagine pouring a bucket of water, but not in 2D, but in 10 million D. Most probably, the water will indeed end up at the lowest point. A puddle, or local minimum, is extremely unlikely.
Why? A puddle means that locally, there is a hill or moat in every direction. In 2D, you just need hills on the X axis to stop the water. In 10 million D, you need a 10 million hills. Since getting a hill in 10 million directions is quite unlikely, you will end up with very few local minima. (See here)
(Note that high D spaces are not without challenges, but I don’t want to get into that now).
So in high dimensional spaces, Moloch wins over Slack. With close to infinite possibilities, you’ll always find a direction to improve the status quo.
Hm, do you feel like the eye example is something that could never happen? What about the example where nobody competes with Boeing because even if they could eventually be better, they would be outcompeted at intermediate stages? If those could happen, how are they different from the machine learning example?
I think the important difference is brought up in Inadequate Equilibria; you need at least two sides that are locked into some bad equilibrium.
The modern ML problems were we don’t worry about local minima are usually only one sided: eg detect cat in a picture. Or, drive a car.
Not: compete with other people for research funding on one side, ans decide on how to hand out research funding on the other side.
While it is true that high-D spaces have very weird properties and many low-D intuitions don’t apply, there is unfortunately an entire field of physics devoted to the fact that in 10 million D, you can get exp(10 million) local minima.
That’s called disordered systems, or glassy physics, and that’s why glass doesn’t flow and also why there are NP-hard problems (including finding Nash equilibria of many-player games for instance).
See Mezard and Montanari’s great book “Information, Physics, and Computation” or this seminar for instance.
The fact that some algorithms might be rescued from this problem doesn’t suggest to me that it is not a relevant picture of the world, very sadly for us all!
I have to admit, I know absolutely nothing about glassy physics. I know that when training neural networks, this issue doesn’t really come up, and there are attempts at theoretically explaining why that isn’t the case.
Do you know how to reconcile these two? Are training NNs a special subset of high D problems where local minima are a non-issue? What am I missing?
Thanks for the refs!
I’ve been going quickly through Sagun et al, Ben Arous is definitely a glassy person so the authors are probably well aware of the issues.
I think the crux is this: under conditions which I need to understand (but may well be general), there is a whole bunch of local minima, but getting into any of them already takes you into a “fairly good” part of parameter space. Not optimal perhaps, but still much better than almost everything you’d get at random.
It’s true that it happens, in spin glasses, that your many local minima have somewhat similar statistics in some respects (though not in others).
In the same spirit, there was a quite controversial evolution experiment on yeast where they deprived it from its food source and replaced it by something inedible.
They saw the population crash, then quickly go back up, and every time they repeated the experiment, the (epigenetic) “solution” found to process the bad resource was different, some definitely better than others, but all allowed you to survive.
Thanks for the help! The more I read about the topic the more confused I get 🙂
I actually encountered the view you wrote in terms of neural network training as well (that there are multiple local minima, but each one is ‘fairly good’).
I think the practical point for not worrying about local minima in NN training is fairly robust (at least for supervised models; GAN training can be fiddly), but I’m not sure what to think of the underlying theory…
Neural networks for supervised learning really seem to be a special case.
In reinforcement learning bad local minima are a real issue. This is fundamentally caused by the exploration-exploitation tradeoff, which does not exist in supervised learning.
In Scott’s terminology, exploration is Slack and exploitation is Moloch. If you have too much of either then you stop making progress, and even with the right amount you are ultimately in the hands of Fortuna.
when training neural networks, this issue doesn’t really come up
What about the effect described in this very interesting paper from Japan? Does it describe a way of avoid a local minimum of the loss function, or a way to find a model with a slightly higher value of the loss function but better generalization? Either way, I think it’s closely related to the problem of avoiding local minima, and I like the way the name “flooding” uses the same water metaphor in a creative way.
Since existing regularizers do not directly aim to avoid zero training loss, they often fail to maintain a moderate level of training loss, ending up with a too small or too large loss. We propose a direct solution called flooding that intentionally prevents further reduction of the training loss when it reaches a reasonably small value, which we call the flooding level. Our approach makes the loss float around the
flooding level by doing mini-batched gradient descent as usual but gradient ascent if the training loss is below the flooding level. With flooding, the model will continue to “random walk” with the same non-zero training loss, and we expect it to drift into an area with a flat loss landscape that leads to better generalization.
The ability of data scientists to manipulate hyperparameters (e.g. regularization parameters) helps resolve the apparent conflicts between physical observations where local optima are inevitable and ML techniques where they are vanishingly rare. Physicists can only observe the fundamental properties of their materials; they cannot change them. It’s feasible to create local optima in a neural network if you do a bad job.
Indeed. Neural network practitioners have developed over the decades an arsenal of tricks to make their models converge to good local minima: residual connections, normalized initializations, normalization layers, adaptive gradient descent with momentum, and so on.
It is true though that just making the model bigger helps. This is essentially because typical neural network architectures become Gaussian processes or something similar in the limit of infinite width, and Gaussian processes have a convex training problem.
Interesting point. This would go in the same direction as William Bialek’s argument (in his book Biophysics), that biological networks within an organism are typically neither glassy nor fine-tuned — understand, neither a landscape with a billion billion minima, nor clockwork with a single well-defined working state and range of mechanical tolerance.
Typical example: useful proteins can fold in a few shapes, not all great. Artificial random polymers of comparable length can have a hundred thousand folding states.
So, at a glance, cell (or organism)-level evolution playing with the hyperparameters of the fitness landscape for protein foldings seems to match what neural network people do.
Never is a strong word. I’m sure there are examples. But how relevant are they practically?
I don’t know the details of the Toyota case, but maybe that fits. Still, is it the rule or the exception? If advising a rising 3rd world dictator on economic policy, would you say that ‘protect your domestic industry & you’ll be rich’ is a good course of action?
So why aren’t people competing with Boeing? Well, there is one startup building supersonic passanger planes, but why aren’t there more?
To be honest, I’m not sure. My guess: Elon Musk could build a better than Boeing company. But we don’t have 1000s of Elon Musks, and the ones that exist work on more pressing problems. I’m skeptical that there are 1000s of Musks whose only obstacle is that they will be competed away at the intermediary stage.
Here is another way to think about it. Under the too-much-Moloch theory, if an asteroid hit the Boeing headquaters, destroyed all equipment, we would be better off (as that would lessen the competitive pressure). I highly doubt that.
The one area where I do see this effect: signaling industries. The diamond business or higher ed / academia. There probably are 1000s of people who could make a ‘startup university’ that’s much better (much cheaper!) than the existing model, but it’s hard for them to get accepted as non-weird (Lambda school may be an exception here).
If an asteroid surgically destroyed all academic institutions, the world may indeed be better off!
They are. Not just Airbus, but also a couple of other countries are trying to break into the passenger jet market.
There are actually two. The problem is that building passenger aircraft is really hard. Not just from an aerospace engineering perspective, but from a regulatory one. I’ve heard that it’s harder to put a part into a human than it is to put (a different) one onto an airliner. I used to be involved with the latter, and the claim is at least plausible.
And when you are competing with Boeing and Airbus, you are not just competing with the companies, but also the USA/France/Germany and their secret services. So you better have your own big country backing you.
Yeah, you need to have solid backing from your country. Bombardier was trying to compete with Boeing and Airbus on medium-range airplanes with the CSeries, which by several accounts was a very promising product. But then Boeing and Airbus both tried to attack the CSeries by cutting prices, lobbying for tariffs, etc., and Canada was probably not big and/or willing enough to really give Bombardier the support they needed. It’s been speculated that it’s because Bombardier doesn’t have a very good reputation outside of Quebec. The government of Quebec did try to help, but was certainly not big enough. In the end, the CSeries was given essentially for free to Airbus and became the A220.
If advising a rising 3rd world dictator on economic policy, would you say that ‘protect your domestic industry & you’ll be rich’ is a good course of action?
See my comment re: Ha-Joon Chang. This was essentially what most advanced economies did – used protectionist policy to grow nascent industries until those industries were powerful enough to compete in a freer market. Samsung and Toyota were specifically protected by their countries in this exact way. The US and UK were very protectionist and only became standard bearers for free markets when they were convinced they would be dominant in those free markets.
See also the Chinese Internet with its Great Firewall, whose primariy purpose was censorship, domestic surveillance and prevention of foreign espionage, but it had the effect of creating a thriving domestic market largely isolated from the American companies. Today China has Internet megacorporations with top-tier technology such as Tencent, Baidu, Alibaba.
Contrast with Europe. There were quite a few promising Internet startups in Europe in the 90s and early 2000s but they have long since folded, now the market is completely dominated by American companies which pay little taxes and employ few people there compared to the profits they make, and of course let the US government access their data as it sees fit.
On the other hand, if you isolate yourself too much (e.g. modern North Korea, or historically, Edo period Japan), you’ll get stuck with obsolete shit-tier technology.
For a country to be “dominant” in one area, other countries have to be dominant in others. There’s no such thing as being dominant in free markets as a whole.
It’s worth noting that some systems, particularly systems operating in physical environments or with developmental constraints, really are constrained to low dimensionality – e.g. lever arms in jaw muscles.
What if you’re designing organisms, and your “fitness” criteria is how close they can get to the Moon? Your algorithm is busy creating birds that can fly to the highest possible altitude. How does your multidimensional optimization get from that to organisms that can build rockets instead?
I’ve trained some neural networks, and my experience is that this is just not the case. Maybe because I’ve mostly been working in reinforcement learning. In that field neural networks get stuck in all kinds of bad local minima that you can’t get out of. Like when playing a video game, the neural network loses points when it dies. And at the beginning of the training, when it is really bad at playing the game, every time that it leaves the first screen it dies to the first enemy. So it learns to just always stay standing still on the first screen because zero points is better than negative points. You have to get really creative to get it out of that local minimum and directly encourage exploration. (oh and you have to start over from scratch. The neural network calcifies in the bad local minimum, and only a random restart can get you to a better place) The same network can represent much better behaviors and actually learn to play the game somewhat well, but it will never get there unless you get really creative while training it.
And, at least in my experience, you find many more local minima than you find ways to get to the global minimum. Most of your work on neural networks is trying to find ways to avoid these countless local minima that you get stuck in.
Fair enough. I was thinking more in terms of supervised learning.
Why not simply reward the network for the number of actions it can complete before it dies, and not penalize it for dying? Staying on the starting screen counts as 0 actions. Leaving the starting screen lets it take some actions. If there’s a game state that lets you move indefinitely without dying or advancing, then add another reward criterion to incent not staying in that state.
Then it learns to take meaningless actions that don’t put it in danger but also don’t get it any closer to actually winning the game.
If the object of the game is to get to the other end of the stage I suppose you could grade it on how close it gets to the other end, though a level design with dead ends would screw that up.
Surely “very few local minima” doesn’t necessarily mean that ending in a puddle is “extremely unlikely”? If you just have 3 local minima, you would naively expect there to be just 1/3 to end up at the global minimum. Since deeper minima are often wider, the probability is probably somewhat greater, but it seems far from guaranteed that you’d end up at the global minimum.
Also, I think it’s extremely rare for big neural networks to actually end up at global minimum. If this was common, you’d expect models to end up with exactly the same behaviour every time you trained them, regardless of the initialization. Do you have an opinion of why this is, if not for local minima? Is it only because we have too little compute?
I think that in many cases, people don’t know the actual global minimum (and why would they?), so if the AI stops improving, they call that the global minimum, while it may just be a local minimum.
For a vivid demonstration of a local minimum that is not a global one in a very-high-dimensional space, consider the giraffe’s recurrent laryngeal nerves.
And you’re deliberately setting them up so that all ten million parameters have a priori equal weighting in the outcome. Reality doesn’t work that way. Reality is, some parameters are more equal than others. Once you’ve optimized for e.g. “Macedonia can support a modestly large army”, “Macedonia has developed a high level of military professionalism”, and “The Macedonian aristocracy occasionally spits out genius-level conquerors”, it hardly matters what the liturgical practices of the Macedonian religion or the growing season of the Macedonian turnip are.
I don’t think think this is true at all.
Rather, any good modern algorithm has considered the problem of local minima and included tools for avoiding getting stuck in them. This doesn’t mean the problem disappeared. It just means that plenty of useful and effective solutions exist.
According to Andrew Ng (one of the pioneers and experts on neural networks), the above commenter is correct – local minima don’t matter with high dimensional models. In sufficiently high dimensional space the odds of a local minima even existing are vanishingly small. Intuitively to get a true local minima there needs to be one along every dimensional axis simultaneously, otherwise you get a saddle point. Saddle points can behave like local minima but there are techniques for getting around them.
See here:
https://arxiv.org/abs/1406.2572
Your argument for why the probability for a particular point being a local minimum sounds convincing, but the number of points, and thus the opportunities for a local minimum, increases massively with the number of dimensions. Also, your argument just as well to *global* minima, and suggests that there are no interior global minima. If we’re constrained to the boundary to find minima, then we’re back to a lower dimensional space.
Btw, you’re still playing Civ 4???
Civilization IV is the best game in the series (fite me). Why play a worse game just because it is newer?
In my opinion, the best game is whichever one my friends are playing, which is why I have Civ V.
It’s also an example of local minima: Civ V is the best one for me because it’s the only one that I know the systems well enough to make intelligent decisions. The others have a >10 hour learning curve to understand that risk/reward pay-offs over the course of a game. In the others, I’m just guessing.
Civ IV does clearly have the greatest music. And the best tech quote readings (outside Alpha Centauri).
I had a similar problem switching from Civilization III to Civilization IV, but Sullla’s tutorial got me over the hillock.
My wife got Civ 6 for the Switch. Neither of us has managed to finish a game.
Our most common Civs are the classic Civ 1, and Civ Revolutions. Lots of people hate Civ Rev, but you can play a full game in one night, which is nice.
+1
I’ve been impressed by Civ VI since the Gathering Storm expansion. V was kind of lame for me, but this one is killing it.
Absolutely. And the best Civ IV mod is Dawn of Civilization.
Can you please explain to me the trade-offs between squares and hexes?
I grew up jealous of my friends who had civ 4, but by the time I got to pick games Civ V was out and hexes just seemed much better.
The problem with Civilization V is not the hexagons, it’s One Unit Per Tile. Based Sullla can explain it much better than I can.
Interesting as I’d have said the one unit per tile thing made the game much better, as it removed doomstacks. I suspect this comes down in the end to people wanting fundamentally different kinds of strategic game experiences. Something like 1upt means that decisions made in an individual turn are far more impactful. Which I prefer because it makes those turns more interesting, as they contain a bigger element of risk. Others would say the opposite because it means a single turn can override a large number of turns of strategic gameplay
*Edit* Occurs to me that the total war games are a more extreme example of this difference. Where good play in the battle element can ovvrride bad play on the campaign map, and vice versa.
Sid Meier’s Civilization is the best, because you can choose the “Play on Earth” option and 4 players (so that the Aztecs are available but the Americans don’t exist) so you get the entire Americas to myself, then amuse yourself by colonizing the continents and rushing tech as fast possible in peace.
As you can see, Scott’s aside brought back a lot of memories for me.
Civ IV was the peak of the Civ formula developed in the earlier games. Most of the changes since then have made it clunkier, more tedious, and less streamlined.
I think this post could use the perspective of “the selfish gene”. Basically, if you think that evolution optimizes individuals you fall into a similar trap as the people who think group-evolution is a thing.
Think of a gene in birds that causes them to throw their siblings out of the nest. In the first generation, such a bird with such a gene would have an advantage (more food from parents), but the bird would have a hard time to reproduce (only one offspring per year), and thus such a gene cannot survive. Exept it does, in cuckoos.
Or take a gene that would cause mice to seek out cats. Exists, in toxoplasmosis.
So, at the bottom level evolution “is about” reproducing genes, and everything else, including other genes on the same chromosome, is just environment. A gene that limits fox population in the abundance of food could only evolve in rabbits.
So, to find an optimal balance between competition and slack, you would need an out-group actor that can create a mutually benifical environment by introducing slack. Investors are a good example. Someone who is already rich cares much more about maintaining wealth long-term than increasing wealth short-term.
In the company context, it would be a CEO that wishes to be famous for reforming a struggling company to be ready for the future, rather than just increasing the ROI by 3%, but who also knows they will find a job somewhere else if the board isn’t happy with their approach.
Not so clear. A lot of rich people got that way by being obsessed with getting richer. They are unsatisfied by their first billion, just as they were unsatisfied by their first million.
Great post (so far), but the example of Italy is… really, really bad.
The Borgias do not represent Italy’s flourishing. They represent the
countryregion wasting away its humongous civilizational and economic advantage acquired in centuries prior. (Wasting it so thoroughly, in fact, that historical GDP reconstructions suggest Italy did not return to its late-middle-ages heights until early XX century.) And Da Vinci and other Renaissance icons were not a product of the ongoing collapse around them. They were a product of slack awarded to them by their filthy rich patrons and their large conspicuous consumption budgets.Not to mention, clocks are wonderful, complicated mechanisms and pinnacles of then-contemporary technology.
It is perhaps worth noting that Welles puts his line in the mouth of a character who has made a fortune by stealing medicine, diluting it to the point of uselessness and reselling it, leading to a hospital ward full of dying children, has committed multiple murders to cover up his activities, will soon be shot dead in a sewer, and is currently making a recruitment pitch for his criminal organisation to the addressee.
Harry Lime is charming and exciting and a wonderful character, but we should probably listen to his advice at least a little sceptically.
There is a very nice (and veeery long) blog series by a historian on why Sparta was different (way worse) than how we imagine it and why they failed to be successful despite starting as the biggest and strongest Greek polis.
The most damning part relevant to this claim is that Spart weren’t really great at war – they won around 50% of battles they fought.
https://acoup.blog/2019/09/20/collections-this-isnt-sparta-part-vi-spartan-battle/
(The blog also explains why the pop-culture perception of Sparta is so different from modern understanding of the historical record)
I’ve written about this before at https://slatestarscratchpad.tumblr.com/post/163974181741/femmenietzsche-tanadrin-spartan-power-as-an
I’d suggest reading the actual post. It talks about how the Spartans appear to have had a slight edge in tactics, but were terrible at every other aspect of war (in particular logistics). Meanwhile the quotes you provide basically seem to be about how badass the Spartans are. But being badass isn’t the same thing as actually being good at war (or even at winning battles).
I read the full series of blog posts, and holy shit, Sparta seems awful. Like, a strong contender for Worst Ever Society. I guess competitors could be flash-in-the-pan badness like the Khmer Rouge and Nazi Germany, but they didn’t last long enough to impose that scale of immiseration. Maoist China and the USSR could compete, and probably come out as causing a greater amount of suffering just because they’re so much bigger.
But as awful as they treated their people, I don’t think they got quite as bad as the overwhelming majority getting hunted for sport, and the relatively lucky few getting taken from their families at age 7, and tortured and raped until they’re ready to continue the cycle. Maybe Belgian Congo?
I am just getting started on the series and holy crap, this is great. Thanks for the link!
Nitpick: You are supposed to win 50% of the battles you fight. Even if the Spartans were the greatest warriors on earth we would expect them to win 50% of their battles. This is since battle is only fought when both sides have an expectation of victory, so the result should be 50:50. If you believe that you’re likely to lose a battle, you don’t fight it: you flee or surrender instead. If the Spartans won 80% of their battles, their enemies would adapt their strategy to avoid the less favorable battles until the percentage went down to 50% again.
Imagine that Spartans were inhuman supersoldiers: that each Spartan could defeat 100 normal Greeks. Now a Spartan army of 100 soldiers march up and is met by an Athenian army of 2000 soldiers. What happens? Well, the Athenians know they don’t have a chance since they need more than 100:1 numbers and only have 20:1, so they flee or surrender without a battle. Ok, let’s say that the Athenians have an army of 20.000 instead. Now the same happens with the Spartans: since they know they need more than 1:100 to win and only have 1:200, they flee or surrender. Battle will only take place when the ratio is about 100:1 and that gives the Spartans even odds of victory.
This assumes that your opponents have an accurate view of your capabilities and that they consider retreat/surrender an option. The absence of those factors is how spectacular generals like Alexander or Napoleon were able to have much better than a 50% battle win-rate.
The post also mentions a factor that may have helped the Spartan’s numbers. The way armies lined up was the most prestigious polis’s phalanx was on the right, and then from right to left in descending order of prestige. So when you have two armies facing each other, who are the supremely-prestigious Spartans going to be lined up against? The least prestigious members of the opposing army, who are going to be fairly intimidated by the Spartan’s reputation.
All true, but A: the Spartans were renowned for their actual military capabilities and B: the battles their opponents chose to fight were in fact the 50:50 ones and C: I don’t recall the Spartans massacring enemies who surrendered or fled before battle(*). So I think all of those conditions apply.
I did toy with the idea of going through the cited battles and plotting the Spartan win rate as a function of numerical superiority/inferiority, but never got around to doing so. That would probably be more indicative of martial superiority than just counting wins.
* The idea that you get to surrender during a battle and live, is a very recent invention and one often ignored by people actually fighting battles.
FWIW I thought the list of battles was the weakest part. What was more compelling to me was that a) the Spartans engaged in pretty much the same sort of fighting as everybody else did, and hoplite battle seems to have been the sort of thing you could pick up with pretty minimal training – there is no evidence the Spartans actually practiced much drill or other combat tactics, b) to the extent they had an advantage, it was more “indoctrination” (unit cohesion) and the fact that Spartiates were probably fitter and better fed than other Greeks who had to work for a living, c) the Spartans May have been better at battles but they seemed to suck at war (really crappy logistics, even for the time, and simplistic strategy)
The bigger nitpicks for me:
1. he includes a bunch of naval battles in his scoring. Sparta didn’t have a navy for a long time, due to being landlocked, and only built one during the Peloponnesian War out of necessity, so of course they weren’t very good at naval combat, since there’s a very definite learning curve involved with all the rowing this way and that in tactical maneuvering to ram enemy ships while they’re trying to ram you (or avoid being rammed).
2. Why does the tale of the tape start at 500 BC? The Spartans conquered the Messenians and reduced them to helot status (thus beginning the era of Spartan militarism) in the 700’s. Seems like one could amass a pretty impressive record in the 200+ years between then and when this chap starts his tabulating. I don’t know enough to say whether that actually happened or not, but still, starting the scoring at 500 BC seems a bit arbitrary.
3. Sparta (along with most of the rest of Greece) was in decline after the Peloponnesian War, so one would expect poorer results after 404 BC, so that’s no surprise. In general, I’m inclined to think ancient Greeks had a clearer picture of who was good at fighting and who wasn’t than people alive today, so I’m skeptical of this kind of revisionist thinking. You’re not gonna get any argument from me with the larger point that Spartan society was not a pleasant one for any strata.
It may have been pleasant (by ancient-world standards) for citizen women on account of when the menfolk were busy oiling up and slaughtering people they left their women mostly alone to run everything else. That’s not the part of Spartan society that everybody glorifies, of course, but I wouldn’t mind a good story told from that perspective (Gorgo’s bit in “300” might be at least a nod in that direction, but I want the version where the women are more pragmatically cynical about the murdery distractions).
The post series covers this. Spartiate women didn’t get to “run things” outside their households (they weren’t allowed in the government). But they did have pretty pleasant lives, because unlike most other Greek women (even relatively noble / well off ones apparently) they were not expected to engage in any sort of productive labor like cweaving, and in fact scorned other Greek women for engaging in those types of activities.
Of course Spartiates were only a tiny fraction of the Spartan population.
1. He addresses this. It doesn’t change the percentages much if you remove the naval battles.
2. He addresses this too. For one thing, we don’t have a lot of actual non-legendary information from this period of Spartan history. For another, smashing up the locals isn’t nothing, but it’s not that impressive because Sparta was significantly larger than all its neighbors. Finally, the authors thesis is not “Sparta sucks at war” it’s “Sparta gained a fearsome reputation early (with a big helping of favorable propaganda via Herodotus) and then coasted on this reputation for a couple centuries where their actual results were middling”
3. But Sparta’s reputation as super tough warriors does not seem to have helped them outlast their neighbors during this decline. If anything Sparta declined a lot harder – Athens stayed relevant much longer, Sparta became a backwater (a literal tourist attraction by the Roman era)
1. I realize that; the Spartans’ record at sea is better than I would have expected, but why include them at all? He spent 18 paragraphs prattling on about phalanxes, then muddies the waters.
2. Okay, but if we don’t have great info, isn’t a more agnostic stance more appropriate? The idea of the Spartans beating up on locals who were apparently small fry seems a bit dismissive, also, because how do you suppose Sparta got to be larger than its neighbors? It didn’t start off that way. The Messenians, if I’m not mistaken, were several times more numerous than the Spartans. And yet, who spent several centuries subservient to whom? Which leads me to another point: not all victories and defeats are equal. Losing at Sphacteria is a bummer, but the Spartans still won the war.
3. That’s fair. That point was mostly aimed towards his comments about how the Spartans didn’t put up a fight against the Macedonians. The jig was largely up at that point, for Sparta, and Macedonia was clearly the New Power to be reckoned with.
On the Sears example, it’s been a while since I read this but I believe P&G were able to find a competitive equilibrium within their own firm. P&G is a massive firm with thousands of different brands, and it is my understanding that they have encouraged their own brands to compete with each other, rather than just those of their competitors, and this was presented as a factor in P&Gs own strength in th FMCG industry.
Unfortunately I can’t remember where I read this so no source, but perhaps there’s an MBA or similar here who has read the same and can back me up
Scott, on Sears, read up on Transaction Costs (Coase). Having departments perform their functions efficiently is great, but they could just be separate firms then. Unless, organizing within a single firm has transaction cost benefits. It’s critical to the decision to do things in-house or outsource. If the benefits are large, having departments strike out on their own is quite literally giving up your organizational advantage. Though to your point, it’s possible that having a temporary period of competition might shake out the cobwebs (unfortunately, this is also the misunderstanding of the business cycle and “creative destruction”). To sohois’s P&G example, it’s about striking the right balance- Tide and Bounty (or however their SBUs are organized) have their own markets to tend to, but for things that they have in common, it’s easier to just ask daddy.
It also relates to your broader point- if you have a bunch of firms transacting with each other, sometimes it might be more efficient (vs. other firms) to organize as a single firm to avoid those transaction costs. If you have a bunch of organisms competing with each other, sometimes it might be more efficient (vs. other groups) to organize as a collective to avoid those competition costs.
There are many companies that have units that sell services internally to other units.
Sears was in an incredibly weak position and retail (especially big box retail) is dying everywhere. Blaming the death of Sears on that particular transformation is a stretch it’s possible that the attempt killed it faster than it would have normally died through many complementary reasons – but blaming internal competition explains entirely too much.
Came here to comment about Coase’s paper as well. Really worth a good read and sad that Coase is more known for The Problem Of Social Cost and not The Nature of the Firm.
+1
I’m not sure that the original “insect-cannibalism” paper entirely supports the conclusions given here or in the lesswrong post.
Specifically when you say they “…killed off any subpopulation whose numbers got too high, and “promoted” any subpopulation that kept its numbers low to better conditions. They hoped the insects would evolve to naturally limit their family size in order to keep their subpopulation alive. Instead, the insects became cannibals…“, this would seem to correspond most closely with “treatment B” in the paper (“group selection for low numbers of adults per population“) which did not have a higher cannibalism rate. Other treatments (the random ones) did produce higher cannibalism rates, and the paper mentions that “some of the B populations enjoy a higher cannibalism rate than the controls while other B populations have a longer mean developmental time or a lower average fecundity relative to the controls”, i.e., sometimes cannibalism, sometimes other factors.
Overall, on a first reading it sounds to me like the conclusion is more “it’s complicated”… but I’m not sure I’ve completely absorbed and understood the paper, maybe I’ve just missed something basic here. I don’t think it makes a big difference to the contents of this particular post but I do think having a clear exposition of experimental support is important, so if anyone with a better understanding of this topic could set me straight that would be useful!
Yeah, this paper is very unconvincing, particularly because their “control” changed so massively, indicating that something else was going on (insufficient nutrients leading to cannibalism?). Add in the small population sizes (prone to drift and inbreeding), and there’s all sorts of trouble. That line B declined in adult population could simply be because the author was unintentionally selecting the messed-up, inbred populations, and what they thought was group selection was just that population circling the mutational drain.
“enjoy a higher cannibalism rate” interesting phrasing
The deep issue with the “group selection problem” is that it sounds a lot like “atoms should generally be found alone, unless they can somehow gain energy by making a molecule”.
That statement is true, but it is also highly misleading as to how many atoms are noble gases. It terribly overstates the unlikelihood of mechanisms that create decidedly collective behavior in chemistry.
Although we (roughly, ecologists and evolutionary biologists) all agree obviously that group selection cannot violate some strict bounds created by individual selection, how often are you close to those bounds? Is it like saying “human language cannot violate the laws of physics”, where really, you’re not looking at the most relevant constraints at all?
I think it is an open and entirely non-trivial question to know whether natural selection gives you mostly noble gases, mostly ocean-sized polymers, 50/50, or something else.
The reasons why the mainstream has alternated between absurdly overemphasizing one or the other are much more sociological than mathematical.
I think there is a conceptual dead end that arises when the statement “evolution doesn’t act on the level of groups” is taken to mean “evolution acts on the level of individuals”. The former may be true, but the latter doesn’t follow. Individuals don’t evolve. Genomes evolve. A gene that kills its host but makes other hosts containing its copies multiply and prosper will win against a gene that prioritizes its current host over anyone else.
That’s just one more turtle on the way down, honestly. You can replace “group and individual” by “organism and gene”, or by “species and group”, or “ecosystem and species”. Or even, if I’m being cheeky, by “gene and atoms”.
All these levels are being selected upon at any given time, the question is “which is the most important to determine current dynamics?”
One naive but helpful way of seeing it is as a matriochka: everything that physics permits > everything that gene selection permits > everything that individual selection permits > everything that group selection permits > …
So, sure, group selection can never violate gene selection or physics. But how close are you to the boundary of each domain? Are you unlikely to evolve a given feature because of thermodynamic bounds, because it would allow selfish genes to take over, because it creates cancer, because it will eventually kill your family or species or planet?
I mean, the amount of biology we can explain without having to resort to group selection has, IMHO, led to the question of whether or not it even matters? If we don’t even need it to explain eusocial insects, it’s either nonexistent, so weak as to be a rounding error, or only occurs in such restrictive circumstances that it’s mostly uninteresting. And the failure to turn up any examples of the latter case, however restrictive or rare, strongly suggests the former two.
I wish I were as optimistic as you are about how much biology we are explaining 😉
My best guess (as someone working in the field, not a naysayer) is that we understand roughly nothing of the main selective forces applying on most species in most ecosystems, and I suspect that it has to do with the fact that we have no idea how to really measure multilevel selection.
4th paragraph: spend -> spent
I have always thought the Greene quote from The Third Man was adapted from Whistler’s 10 O’Clock Lecture:
– but there the Swiss are used as an example of the opposite of too much slack, as people accustomed to fighting for their freedom (Gessler, Tell), who are less favored by art than are the languid opium eaters.
I’ve been thinking about the Japanese car industry, and how perhaps the biggest protectionist step the Japanese government took was to have traffic drive on the left (which obviously discourages importing cars from the US with their steering wheel on the left!).
I’m also looking at the UK and Continental Europe as comparisons. In both cases there were non-tariff barriers that made importing cars directly from the US difficult. The most important one in the UK was left-hand traffic (as in Japan), but there were also factors all over Europe that meant that people preferred smaller cars with smaller engines- they fit better in narrow streets, fuel is more expensive, and many countries had high taxes on heavier or more powerful cars.
The response of Ford and GM was to have subsidiaries in Europe producing vehicles aimed at the European market, largely using European-produced parts. They went about it differently- Ford set up Ford of Britain and Ford of Germany, while GM took over pre-existing manufacturers Vauxhall and Opel- but the end result was pretty much the same.
I was wondering why no American companies tried this in Japan. Might it just be an issue of timing?