The Adjacent Possible: Steam Engines to Neural Nets

Recall our Chapter 7 discussion of accelerating societal change as a combinatorial explosion due to all the possible combinatorial ways to innovate when new multipurpose physical tools (eg, smartphones) and informational algorithms (eg, deep learning) become widely available, and are not just restricted to elites.

To better understand how portals emerge in an adaptive landscape, complexity theorist Stuart Kauffman brings us another key concept, the adjacent possible, which describes the first-order possibilities for combinatorial innovation that are always opening up in the local environment at any particular time. Computationalists and physicists might call this the new “phase space” of obvious next combinations of ideas and tools that are experiments waiting to be tried out.

Science writer Steven Johnson in his lovely Where Good Ideas Come From: The Natural History of Innovation (2010), uses Kauffman’s idea of the adjacent possible to explain what he calls “multiples”, the convergent parallel invention or discovery of important new scientific and technological ideas. Kevin Kelly also studies convergent evolutionary development in his epic look at the history and meaning of technology, What Technology Wants (2010). Both show how incredibly common convergent innovation has been over the centuries. An early study by Ogburn and Thomas, Are Inventions Inevitable? (1922) found 148 instances of independent convergent scientific discovery or technological innovation, most within the same decade.

As Johnson might say, in communities and cultures of a roughly similar idea and technology complexity level, and with reasonable information flow between cultures, all innovators will be working in the same mental space of the adjacent possible. That means they’ll all be primed to most easily see those potentially promising first order combinations of new ideas, theories, and tools. When we understand that the envelope of adjacent possibles is a constraint on the innovation environment we understand how innovators will often arrive at the same adjacent possibles (experiments) and adjacent optimals (discoveries) at similar times.

Studying two hundred of the most important innovations since Gutenberg, Johnson further notes that what motivates (evolutionary) experiments, far more than either the genius of lone individuals or the profit motive, is the density, diversity, and freedom of the innovation networks in which new ideas emerge. Per Chapter 7, we can say the more dematerialized (freedom, diversity, and information-enabled) and densified (STEM-dense and STEM-efficient) our innovation networks are, the faster great experiments and great discoveries will occur.

Taking an evo devo approach, we can recognize that innovation convergences happen in two basic types, experiments and discoveries. An experiment is something that may have some limited social success in a community (scientific, technological, political, social), but which doesn’t turn out to be particularly universal or global. A discovery, whether in science (discovery of oxygen), technology (the wheel, the digital computer) or politics (democracy, universal human rights) is something that becomes optimal in many contexts, for very long periods of time. So innovations can be either experiments or discoveries, and the latter are far rarer than the former.

Let’s look at an example. In his vocational advice book, So Good They Can’t Ignore You (2012) computer scientist Cal Newport relates that while attending a professional conference in 2011, he counted four different presentations, by professors at different universities, each applying a previously obscure computational technique, randomized linear network coding, to the problem of information dissemination in networks. Whether this convergent experimentation becomes a convergent discovery or not depends on how broadly useful and universal this tool turns out to be. Is linear network coding so optimal for certain computational problems that it would be arrived at by other cultures, or on other Earth-like planets, even if they took a very different path to it? If so, it is truly developmental. The better our simulation gets, the better we can identify and discriminate between experimentation portals, which persist for a time, aided by imitation, communication, or convergent thinking, and development portals, which are far more long-lasting and universally optimal.

If the 95/5 rule applies, we should expect that the vast majority, or roughly 95%, of adjacent innovation convergences are going to be evolutionary experiments (adjacent possibles), innovations that will have a bit of success and be contingently adaptive, for a time, among other near-equally adaptive options. A critical subset, perhaps 5% or less, of these innovation convergences will be developmental convergences (adjacent optimals), broadly and globally adaptive new innovations that will dominate their niches for their time and place, once they arrive.

It’s also important to realize that as combinatorial explosions of experimental possibilities grow, we will devise new intelligence methods of pruning, integrating, and synchronizing those possibilities. This is another form of development. For example, computer scientist Aaron Sloman in “Why Robots Will Have Emotions” (1981) argues that emotions are an integrator and decider that we,  and all future AIs, must increasingly rely upon to make decisions, when we face a combinatorial explosion of possible reasoning threads, and can’t decide via pure rationality in any reasonable timeframe. Consciousness itself is another such integrator, giving us a way of pruning all our subconscious thinking states into one synchronized workspace, and making decisions with a vastly reduced, synchronized, and integrated subset of information.

To summarize these observations, recalling the Five Goals and Ten Values, not only do combinatorial explosions lead to new forms of evolutionary experimentation, via increasing social and technical information, innovation, and indeterminacy, but they also lead to new forms of developmental convergence, via increasing social and technical interdependence, immunity, and inertia. Finally, they lead to new forms of intelligence, better ways of integrating, virtualizing, and deciding.

So when we scan the adjacent possible, and see portals that we think will advance the Ten Values in positive ways, we are practicing adaptive management. Again, we can ask: Is this kind of evo devo thinking in line with “What the Universe Wants,” to riff on Kevin Kelly’s great book, What Technology Wants (2011)? Is evo devo thinking, the Five Goals, and the Ten Values a better way to view the past, to make decisions in the present, and to imagine the future? Perhaps only the future can decide. In the meantime, let’s explore a bit more, so you can better decide how useful it is yourself as a foresight practitioner today.

As we’ve seen earlier in this chapter, one of the most famous and accelerative (runaway) portals, steam power, was at the heart of the industrial revolution. Steam power, the use of carbon fuels to boil water, and devices to turn that boiling energy into work, was first used in toy Aeolipiles beginning in 200 BCE, as described by Hero of Alexandria in Roman Egypt. But social conditions weren’t right for practical steam engines to emerge. Finally in 1712, nineteen centuries later, Thomas Newcomen’s piston steam engine pumped the first water out of a coal mine, setting us up for the industrial revolution. For the first time (excepting wind and water mills, which can be more powerful but aren’t so small and portable) we had dense, portable engines vastly stronger than human or animal muscle, useful for all sorts of productive purposes.

Steam engines (external combustion engines) burning carbon fuels (wood, coal) are an obvious portal for civilization development, and also the most direct route to both internal combustion engines, and dynamos and electric engines, additional accelerative portals waiting just beyond steam engines, on any planet with water and carbon fuels.

Now imagine how different humanity’s past would have been if we’d had philosophies, ideologies or even religions emerge which led us to believe such portal paths existed. We’ve had glimmerings of such perspectives in the past, in war and other crises. For example, we saw this philosophy in the Depression-era Technocracy movement in the US, in Stalin’s Russia and Mao’s Great Leap Forward, but there it was headed by top-down autocrats and utopian dreamers who didn’t understand that freedom, democracy, incentivizing policy, industrial capitalism and entrepreneurship have always been the best engines of technological progress since its emergence in the modern West. How can we validate such a claim? By paying attention to D&D and the Seven I’s. Does a portal truly advance them, in a way leading to greater adaptiveness?

If we’d had sufficient entrepreneurial social capitalism, scientific investigation, and technological innovation ideals emerge and persist in any culture, at any time in the nineteen centuries between the Aeolipile in 200 BCE, and Newcomen’s steam engine in 1712, we could have discovered practical steam engines, and jump-started the first industrial revolution far sooner. I would claim we would inevitably be living in a far more developed and socially mature world today. We would have had the growing pains of oligarchic and plutocratic industrial capitalism far earlier, and would be much further along in implementing their solutions.

Just like our Printing Press in Ancient Greece counterfactual in Chapter 1, this Steam Engine in the Roman Empire counterfactual argues that there are objectively better futures waiting for us all the time. We just need to get better at looking for them, and extrapolating their consequences. Would a commercially successful steam engine in 200 BCE have been truly better for civilization, in terms of the Five Goals and Ten Values?

Such an engine would have been a hard sell in a world with plentiful slaves, and an Empire suspicious of technology, but it could still have found use in aqueducts, home plumbing, flour mills, woodcutting, stonegrinding, and some other killer apps. Let’s leave that question for your consideration for now, and explore it in a later version of the Guide.

This kind of thinking helps us imagine that we’ve got a handful of major portal paths lying right ahead of us right now, waiting for us to wake up and see them. Remember Biologically-Inspired Computing? As we discussed in Chapter 7, perhaps the most exciting portal of all is the copying of neural, evolutionary, and developmental algorithms into our machines, allowing them to run VCRIS cycles and quickly grow to our level of pattern recognition, and then far beyond.

We’ve seen fits and starts of works on this project since the 1930s, and the field has seen a resurgence since 2009 with the emergence of deep learning on parallel GPUs and with a handful of efforts at neuromorphic chips, in particular Dharmendra Modha’s Synapse chip work IBM Almaden, and Narayan Srinivasa’s STDP neuromorphic chips at HRL. But we spend a pittance on this kind of work as a species today, and on the neuroscience that would expose more of these magic algorithms.

There is a lot of unclaimed wisdom in our past, but we humans often prefer to dwell in nostalgia for our primitive ways. Nevertheless, the portal calls, pulling ever more of us to its gates every year now. We can’t stop our species going through, all we can do is determine when and how we will do so. We hope this Guide, and the programs that spring from it in coming years, will help any who read it to walk a better path.