On the days when Alex worked alone at the Middleton Theater, he sold customers their 99¢ tickets outside and their Mountain Dews inside, and then scrambled up to the projection booth to start the movie. Up in the booth, the film ran from one enormous rotating platter, through rollers and across the projection lens, to an identical receiving platter that wound it inside out for the next run. One day, after Alex had started Free Willy and was back in the lobby dusting the Good & Plenty boxes, a customer showed up to say the screen had been blank for ten minutes. Rushing upstairs, Alex found the platter pouring film onto a hopeless tangle on the floor. In a panic, he cut the entire tangle out of the film and respliced it, and for the rest of its run Free Willy began ten minutes into the movie. Although the customers didn’t seem to care, they surely noticed the abrupt jump from the middle of a local advertisement for a “Murcree” car dealership to a scene of Willy the orca swimming around.
The change was sudden and illustrates two points about the future of cultural evolution. First, a gap in an expected narrative upsets our working memory, which is an essential aspect of cultural capacity, and becomes an issue as algorithms and artificial intelligence take on more prominent roles in cultural transmission. Chimpanzees have enough working memory to make complex tools and maintain rudimentary behavioral traditions, such as remembering how to use a tool they haven’t seen in years, but humans can use memory for much, much more, such as remembering subprocedures embedded within larger sequences that are themselves embedded in cultural memory. People can carry out complex activities on timescales ranging from a few seconds, such as solving simple algebraic problems, to many years, such as raising children.
Second, feeling disjointed, as with moviegoers watching an ad for a Mercury car dealer in one moment and then an orca swimming around in the next, has become our way of life. As Thomas Friedman put it in Thank You for Being Late, technology may already be changing faster than human behaviors, laws, institutions, and customs can adapt. This isn’t generational change, as Alvin Toffler described in Future Shock, but rather intragenerational change. It’s happening in all three elements of cultural evolution—variation, transmission, and sorting—through media that are both diverse and rapidly changing. A 2014 survey of US teenagers, for example, still ranked Facebook as the national favorite social medium, but it was already challenged by Instagram, Snapchat, Vine, Tumblr, and newer platforms. This doesn’t count all the messaging apps such as WhatsApp, Viber, and many more that collectively have more users than those big social media networks. Rising and falling in popularity, each new social media or messaging platform imposes its own unique biases on variation, transmission, and sorting, all leaning toward a company’s goals as much as the user’s.
If we take a step back, this unsettling flux represents a transition between modes of cultural evolution. We’d like to use an analogy that is helpful, if not taken too literally. Let’s think of the memory component of transmission as the depth of water in an ocean—shallow in places and deep in others. Our ocean has three different kinds of animals inhabiting it, each representing a mode of cultural evolution past, present, and future. Bluefin tuna are in the deep end, massive schools of herring are in the shallow end, and swimming throughout the ocean we have orcas. Let’s take a closer look at our ocean dwellers.
Bluefins represent local traditions stretching back deep in time. They can dive down to a kilometer or so, they travel in relatively small schools, and they can collectively remember things such as distant migratory locations. Also, like traditional cultures, bluefins are in danger of disappearing. Herring, on the other hand, move in massive schools—as large as millions of fish—and spend time in shallower, coastal waters. Similarly, within a shallow time depth, algorithms guide human followers like schools of herring, using popularity as a beacon. Tank experiments show how a robotic fish moving unwaveringly in one direction can lead a school of real fish with it, just as we saw in chapter 1 with Ian Couzin’s birds. Primates can be led the same way: a few baboons can lead the whole troop to a new food patch if those few travel in the same direction. Animal scientists call it directional agreement. Marketers want the same result and employ data-mining companies to amass thousands of data points per person, aiming for feedback among targeted advertising, human response, and more finely targeted advertising to lead consumers in a certain direction.
Clearly, the more that algorithms facilitate human communication and decisions, the more profoundly they change both the tempo and mode of cultural evolution. For online choices, a popularity bias is driven by search algorithms, which effectively rank options by popularity or network centrality. In social media, positive ratings are a prime currency. Popularity increasingly outranks quality. Whether it’s a hotel room, an investment, or even a scientific theory, people are more likely to endorse something once others have endorsed it. Sure, quality is part of a rating, but if people are copying each other’s errors—which can be on a magnitude scale, where we think a quantity is in the hundreds when it is really in the thousands—the errors do not cancel out. Rather, they feed back multiplicatively into the crowdsourcing algorithms. In 2013, for instance, Google Flu overestimated its influenza estimate—which assumes people Google flu-related terms from their direct experience—because many people were Googling what other people were Googling, which in turn led Google to suggest those search terms, and so on.
In cultural evolution, a shallow time depth means freedom from the deeper past, which often allows more turnover and drift. In 1960, David was the most popular name for boys in virtually all states west of the Mississippi River, with Michael, James, Robert, and John rounding out the top five in almost every state. Now, baby names in the United States are freely chosen—no longer traditional or inherited—and the invention rate has tripled in the last several decades. Turnover in the top hundred is rapid. Sublists guide parents to just the right name for their social group or region—like Addison and Beulah among the best southern names. This has balkanized the landscape of naming. By 2015, neither of the top two boys’ names in Wisconsin—Oliver and Owen—was among the top thirty in California. This is typical of ecological drift. Geographic dialects, for example, evolve among the songs of birds, which copy each other with changes arising through recombination, invention, or errors.
Likewise, drift in social media content creates polarized groups. The silos of fake news on social media suggest linked, drifting ideas that bundle into identities. A “vast satellite system” of fake news sites now surrounds mainstream media sites, as Jonathan Albright of Elon University described it. Algorithms help this happen because the network of fake news sites is highly entangled, such that each site already is well connected, which helps it climb up Google’s PageRank algorithm, thereby prioritizing sites by network connectivity and popularity rather than by validity.
Similarly, social media and crowdsourcing may be making schools of scientific thought more herring-like. Swimming in shallower waters, in terms of more recent scientific bibliographies, scientists are increasingly crowdsourcing their attention through social media feeds. Algorithms used by Google Scholar, Mendeley, ResearchGate, and Scizzle feed articles to scientists through a personalized balance between the scientist’s topical interests and their social as well as citation networks. Other scientists code their own Twitterbots to automatically scan for articles with specialized keywords—in the process attracting hundreds of scientists as followers.
To counterbalance this, the journal Nature advised scientists to “go to seminars and meetings,” and it quoted a young scientist who opined, “Weekly events can help bring people out of their offices, [and] create a sense of community.” If it seems remarkable that anyone would need to be reminded of this, that’s because science has already moved decisively toward virtual collaboration among scientists, their algorithms, and the entire scientific record.
Orcas are perfect models for future knowledge grabbers: intelligent, selective hunters that, working individually or collaboratively, choose their prey from anywhere and at any depth down to their absolute limit. Let’s see how these qualities might apply to scientists. Ideally, open, online collaboration would move scientists toward building on the latest and most relevant science and away from schooling, like herrings, around citation statistics and network links. An encouraging model is GitHub, the Wikipedia of software design, where thousands of developers openly collaborate on projects. Fueled mainly by validating comments from their peers (“Good job!”), the GitHub community dives deeply and even does free projects for giants such as Microsoft and Hewlett-Packard. On GitHub, “ultimately you have an expert—the person who wrote the original program,” noted Thomas Friedman, “who gets to decide what to accept and what to reject.” With the best ideas vetted by experts, GitHub exemplifies the Tasmania model we looked at in chapter 9, in which progress is accelerated by a large population size. The massive global population of orca-like GitHub developers can complete projects in a fraction of the time that it might take a team of paid herring-like employees.
As our allegorical orcas dive deep for prey, they encounter millions of scientific articles lying in their cold underwater tomb. As we saw in chapter 9, many of those articles, especially the older ones, have never even been cited. Virtually any article, no matter how obscure or old, can turn up in an orca’s search. “Nothing in the past is lost. … [E]verything exists on one plane,” wrote poet and journalist Dan Chiasson, so if you are a researcher with brilliant, but uncited articles, take heart, because it’s a good bet that they will one day be brought to the surface. Drawing on digital information both globally and historically, knowledge evolution will be aided by artificial intelligence to select for well-specified qualities. If open science adopts this kind of expert selection, science will shift up a gear. Following the release of the gene-editing technology CRISPR in 2014, Kevin Esvelt of MIT saw open science as morally imperative, especially as humans begin engineering the evolution of animals, insects, plants, microorganisms, and possibly even themselves.
A fully open science can also study itself in order to optimize its own evolution. Forecasting future citations of new medical papers, for example, can help predict whether a drug approved by the Food and Drug Administration might appear a decade later. Deeper insights will come through meta-analyses of text-mined scientific publications. Biomedical researchers at the University of Manchester, for instance, are looking for disciplinary trends and mapping the flow of information between them. Identifying such networks can highlight key new areas of research—potentially those that an algorithm could use to generate a new hypothesis. “Algorithms will build on algorithms,” promised a trade magazine at Hewlett-Packard, “with every prediction smarter than the last.” Predictive algorithms learn through a process called supervised learning, where thousands of successive estimates are checked against the correct answer, each time adjusting the model parameters after each trial to improve the estimate incrementally. This is the essence of Bayesian modeling.
The algorithmic approach needn’t be restricted to published literature; it can also study real people. It could use a platform such as Amazon’s Mechanical Turk, which now hosts over twenty thousand online participants per month on experiments ranging from rating facial attractiveness to studies of generosity and religiosity. Online social research already has global reach, as over half the world’s population has mobile phones. Machine learning can infer much from basic phone data, even the personal wealth of someone in a developing country. Researchers recently compared billions of interactions on Rwanda’s largest mobile phone network to personal phone surveys that provided direct estimates of personal wealth. Machine learning estimated the personal wealth from the phone contacts, volume, and timing of calls or texts and geolocations. It could even predict whether a person owned a motorcycle or had electricity in the house.
Anonymous phone data could also be used to predict conflicts. No conversation content is needed. Instead, all you need is the timing of events, as they tend to accelerate in a predictable pattern, like a ball dropped on the floor: bop, … bop, … bop, bop, bopitybopitybopbopbop. As a conflict escalates, the shortening time intervals between responding events—whether years, days, or seconds—are inversely proportional to the numerical order of the event, via a negative exponent called the escalation parameter. The method has been developed on data sets on the ground and online, regarding escalations of warfare as well as online discussions preceding an attack or civil unrest. It even applies to a fight at the family dinner table (you can verify this with a stopwatch and Will Ferrell’s classic “I Drive a Dodge Stratus!” skit on YouTube).
After wealth and conflict, the next step is to predict health. The future promises many marvels. Hewlett-Packard says that by 2030, your embedded microchips will alert you when it’s time to 3-D-print yourself a new kidney. This is not so farfetched. Even now, a Google user’s search activity—for certain diagnostic symptoms of, say, the onset of diabetes or a chronic condition—can reveal a developing health problem before it is even known to the user. Governments are interested too, of course, at the public scale. In the United Kingdom in 2015, the National Health Service agreed to share millions of personal health records with DeepMind, the Google-owned company that developed a neural network called a “differentiable neural computer” that learns to understand narratives, analyze networks, and solve complex logistical problems.
Speaking of which, did we mention that the orca has a huge brain? Neural networks aim to solve problems the same way a human brain would, with layers of networks that discover patterns of patterns. An image of a face might enter the input data layer, which is passed through layers of intermediate representations, like the edges that make a shape, and then the shapes that make a face, to the response layer. A dynamic neural network learns by rewiring its neurons to reinforce whichever millions of neurons that activated—“voted”—for the correct answer. Given that the number of connections in a neural network grows with the square of the number of nodes, the pattern recognition of neural networks can become much more granular real fast.
Nevertheless, we have not yet built machines that truly reason like people. Researchers at MIT’s Center for Brains, Minds, and Machines say artificial intelligence needs to move from mere pattern recognition—no matter how sophisticated or fast—to causal explanation, which we covered in chapter 8. Many of the advances in artificial intelligence have come about through playing games, which requires a great deal of supervised trial-and-error learning with feedback about the correct answer. To beat a player at Go, for example—even one who’s new to the game—a deep neural network must first observe millions of moves by expert players and play millions of practice games. Facebook’s deep convolutional network needs thousands of examples to judge how a tower of just several toy blocks will fall, which is something a child would know intuitively.
In game-playing terms, artificial intelligence is still chess-like, trained to optimize the long-term reward of a particular action in a particular situation. It struggles to interpret novel input, such as reading a new style of handwriting. It cannot easily generalize or combine simple elements into complex concepts with infinite possibility—a feature of human thought and language known as compositionality. To make conversation, a neural network predicts the next sentence based on the previous one. In this sense it recalls the 1960s’ MIT “Eliza” program, which faked its way through a conversation in “phrases tacked together like the sections of a prefabricated henhouse,” as George Orwell once described the language of politicians. A half century later, the neural network shows more originality. “What is immoral?” Google researchers asked their neural conversational machine. “The fact that you have a child,” it replied, somewhat ominously. To be fair, it and other intelligent personal assistants like Amazon’s Alexa are designed to answer customer service questions or sell products, not to make original conversation compositionally or develop causal explanations about the world.
Compared to present artificial intelligence—distant future versions may be reading this and “laughing”—humans learn a lot more from much less. Learning both individually and socially, children can isolate variables and test causal hypotheses. Children who are taught how to learn, such as at Montessori schools, acquire measurable advantages in language, math, creativity, social interaction, and understanding. Humans can generalize explanatory concepts from just a few examples. To get closer to human creativity and flexibility of reasoning, artificial intelligence must become compositional and thus be able to generalize rather than simply look up each answer from an encyclopedic reference set.
This is precisely the goal of researchers who are experimenting with stochastic programs that can parse objects and goals into their essential components and then recombine them into new concepts and larger goals. This brings us back to the importance of memory. A breakthrough for DeepMind’s neural computer came about through the integration of external read-write memory with the powerful neural network. This allows the computer to represent and manipulate complex data structures, and like a neural network, learn from the data. Just as humans have better working memory than chimpanzees, memory may be the key to humanlike artificial intelligence. Science fiction knows this. At the end of 2001: A Space Odyssey, Hal becomes less human as his memory is unplugged. In HBO’s Westworld, robots acquire human reasoning through their growing retention of personal memories.
As memory and artificial intelligence are integrated, however, and searchable digital records bring everything to one plane, we need to remember how to forget, as Elvis sang in 1955. Our metaphorical orca is not a fish but rather a mammal, and it occasionally comes up for air and to clear its mind. At the population scale, forgetting re-sorts existing variation, cleans the slate, and starts a new phylogenetic branch. We may owe our cultural modernity to this. About seventy-five thousand years ago, the volcanic eruption of Toba, in Sumatra, blanketing southern Asia in ash, might have left fewer than ten thousand people on the planet. Some paleoanthropologists believe the Upper Paleolithic era emerged out of Toba’s ashes, with art, modern behaviors, and new technologies, all of which form the cultural foundation of humanity.
This is a pertinent question as memory and artificial intelligence become more integrated—literally, as machines begin to make decisions. Neuroadaptive technology already exists that can learn to interpret simple human intentions directly from brain activity. While a person moves a cursor on the screen of a computer that is simultaneously doing real-time analysis of the person’s brain activity at five hundred hertz, through dozens of electrodes configured around the scalp, the neuroadaptive system learns through trial and error how to translate the brain activity directly into the intended movement of the cursor. In short, the computer literally reads the person’s mind. As with artificial intelligence in general, however, the question is, How big is the gap—in this case, between simple intentions, such as moving a cursor, and real thought and causal explanation?
In closing, we’re reminded that evolution’s direction is just too unpredictable to predict its future. As Walt Disney said in 1954 on What Is Disneyland, “I only hope we never lose sight of one thing: It was all started by a mouse.” Steve Jobs might have said the same thing, except that Xerox had tried it before him. Remember the Xerox 8010? We didn’t think so. The point is, no one could have guessed the trajectories that entertainment and the computer industry would take after those mice came along. They were products of countless cartoon characters and technological devices that came before them, and which one changed everything is only clear in hindsight.
One thing can be predicted with certainty, though: it’s impossible to know exactly what cultural game changer is on the horizon. The best you can do is try to survey the pool of variation, but that’s a tall order—an impossible one, actually. Our suggestion, not surprising, is to start with technology. For an entry point, you might browse the annual Consumer Electronics Show (now CES), which bills itself as “the launch pad for new innovation and technology that has changed the world.” Just a few years ago, CES was awash with gadgets that connect to the Internet (and each other)—the “Internet of Things”—but in 2017, the show was, as Forbes noted, all “about making more things that create and use intelligence.”
This leads us to a final question: Will technological and cultural evolution continue accelerating indefinitely, or is there a terminal velocity—a point of resistance that causes a “deceleration”? Already in 2017, for example, Facebook and European governments plan measures to curtail the spread of fake news, and surely artificial intelligence will be used—a case where selection is being bumped up in its balance with variation and transmission. As we wonder whether people one day will marry their intelligent, loving artificially intelligent assistant, like in the movie Her, we can also think about general evolutionary processes. To anticipate the future of cultural evolution, think about populations, not individuals, and certainly not yourself. How will variation, transmission, and selection be affected? What feedbacks will arise or be eliminated? Rather than latch onto a single prediction about what the future of culture holds, think like a Bayesian: How does this change the landscape of probabilities? What wave should we be surfing now, and how will we find the next wave after that? Willy and his orca friends call this fun, and you should, too.