Learning, evolution and the future of work

The Janus-headed rise of AI has prompted many discussions about the future of work.  Most, if not all, are about AI-driven automation and its consequences for various professions. We are warned to prepare for this change by developing skills that cannot be easily “learnt” by machines.  This sounds reasonable at first, but less so on reflection: if skills that were thought to be uniquely human less than a decade ago can now be done, at least partially, by machines, there is no guarantee that any specific skill one chooses to develop will remain automation-proof in the medium-term future.

This begs the question as to what we can do, as individuals, to prepare for a machine-centric workplace. In this post I offer a perspective on this question based on Gregory Bateson’s writings as well as  my work and teaching experiences.

Levels of learning

Given that humans are notoriously poor at predicting the future, it should be clear hitching one’s professional wagon to a specific set of skills is not a good strategy. Learning a set of skills may pay off in the short term, but it is unlikely to work in the long run.

So what can one do to prepare for an ambiguous and essentially unpredictable future?

To answer this question, we need to delve into an important, yet oft-overlooked aspect of learning.

A key characteristic of learning is that it is driven by trial and error.  To be sure, intelligence may help winnow out poor choices at some stages of the process, but one cannot eliminate error entirely. Indeed, it is not desirable to do so because error is essential for that “aha” instant that precedes insight.  Learning therefore has a stochastic element: the specific sequence of trial and error followed by an individual is unpredictable and likely to be unique. This is why everyone learns differently: the mental model I build of a concept is likely to be different from yours.

In a paper entitled, The Logical Categories of Learning and Communication, Bateson noted that the stochastic nature of learning has an interesting consequence. As he notes:

If we accept the overall notion that all learning is in some degree stochastic (i.e., contains components of “trial and error”), it follows that an ordering of the processes of learning can be built upon a hierarchic classification of the types of error which are to be corrected in the various learning processes.

Let’s unpack this claim by looking at his proposed classification:

Zero order learning –    Zero order learning refers to situations in which a given stimulus (or question) results in the same response (or answer) every time. Any instinctive behaviour – such as a reflex response on touching a hot kettle – is an example of zero order learning.  Such learning is hard-wired in the learner, who responds with the “correct” option to a fixed stimulus every single time. Since the response does not change with time, the process is not subject to trial and error.

First order learning (Learning I) –  Learning I is where an individual learns to select a correct option from a set of similar elements. It involves a specific kind of trial and error that is best explained through a couple of examples. The  canonical example of Learning I is memorization: Johnny recognises the letter “A” because he has learnt to distinguish it from the 25 other similar possibilities. Another example is Pavlovian conditioning wherein the subject’s response is altered by training: a dog that initially salivates only when it smells food is trained, by repetition, to salivate when it hears the bell.

A key characteristic of Learning I is that the individual learns to select the correct response from a set of comparable possibilities – comparable because the possibilities are of the same type (e.g. pick a letter from the set of alphabets). Consequently, first order learning  cannot lead to a qualitative change in the learner’s response. Much of traditional school and university teaching is geared toward first order learning: students are taught to develop the “correct” understanding of concepts and techniques via a repetition-based process of trial and error.

As an aside, note that much of what goes under the banner of machine learning and AI can be also classed as first order learning.

Second order learning (Learning II) –  Second order learning involves a qualitative change in the learner’s response to a given situation. Typically, this occurs when a learner sees a familiar problem situation in a completely new light, thus opening up new possibilities for solutions.  Learning II therefore necessitates a higher order of trial and error, one that is beyond the ken of machines, at least at this point in time.

Complex organisational problems, such as determining a business strategy, require a second order approach because they cannot be precisely defined and therefore lack an objectively correct solution. Echoing Horst Rittel, solutions to such problems are not true or false, but better or worse.

Much of the teaching that goes on in schools and universities hinders second order learning because it implicitly conditions learners to frame problems in ways that make them amenable to familiar techniques. However, as Russell Ackoff noted, “outside of school, problems are seldom given; they have to be taken, extracted from complex situations…”   Two  aspects of this perceptive statement bear further consideration. Firstly, to extract a problem from a situation one has to appreciate or make sense of  the situation.  Secondly,  once the problem is framed, one may find that solving it requires skills that one does not possess. I expand on the implications of these points in the following two sections.

Sensemaking and second order learning

In an earlier piece, I described sensemaking as the art of collaborative problem formulation. There are a huge variety of sensemaking approaches, the gamestorming site describes many of them in detail.   Most of these are aimed at exploring a problem space by harnessing the collective knowledge of a group of people who have diverse, even conflicting, perspectives on the issue at hand.  The greater the diversity, the more complete the exploration of the problem space.

Sensemaking techniques help in elucidating the context in which a problem lives. This refers to the the problem’s environment, and in particular the constraints that the environment imposes on potential solutions to the problem.  As Bateson puts it, context is “a collective term for all those events which tell an organism among what set of alternatives [it] must make [its] next choice.”  But this begs the question as to how these alternatives are to be determined.  The question cannot be answered directly because it depends on the specifics of the environment in which the problem lives.  Surfacing these by asking the right questions is the task of sensemaking.

As a simple example, if I request you to help me formulate a business strategy, you are likely to begin by asking me a number of questions such as:

• What kind of business are you in?
• Who are your customers?
• What’s the competitive landscape?
• …and so on

Answers to these questions fill out the context in which the business operates, thus making it possible to formulate a meaningful strategy.

It is important to note that context rarely remains static, it evolves in time. Indeed, many companies faded away because they failed to appreciate changes in their business context:  Kodak is a well-known example, there are many more.  So organisations must evolve too. However, it is a mistake to think of an organisation and its environment as evolving independently, the two always evolve together.  Such co-evolution is as true of natural systems as it is of social ones. As Bateson tells us:

…the evolution of the horse from Eohippus was not a one-sided adjustment to life on grassy plains. Surely the grassy plains themselves evolved [on the same footing] with the evolution of the teeth and hooves of the horses and other ungulates. Turf was the evolving response of the vegetation to the evolution of the horse. It is the context which evolves.

Indeed, one can think of evolution by natural selection as a process by which nature learns (in a second-order sense).

The foregoing discussion points to another problem with traditional approaches to education: we are implicitly taught that problems once solved, stay solved. It is seldom so in real life because, as we have noted, the environment evolves even if the organisation remains static. In the worst case scenario (which happens often enough) the organisation will die if it does not adapt appropriately to changes in its environment. If this is true, then it seems that second-order learning is important not just for individuals but for organisations as a whole. This harks back to notion of the notion of the learning organisation, developed and evangelized by Peter Senge in the early 90s. A learning organisation is one that continually adapts itself to a changing environment. As one might imagine, it is an ideal that is difficult to achieve in practice. Indeed, attempts to create learning organisations have often ended up with paradoxical outcomes.  In view of this it seems more practical for organisations to focus on developing what one might call  learning individuals – people who are capable of adapting to changes in their environment by continual learning. 

Learning to learn

Cliches aside, the modern workplace is characterised by rapid, technology-driven change. It is difficult for an  individual to keep up because one has to:

• • Figure out which changes are significant and therefore worth responding to.
  • Be capable of responding to them meaningfully.

The media hype about the sexiest job of the 21st century and the like further fuel the fear of obsolescence.  One feels an overwhelming pressure to do something. The old adage about combating fear with action holds true: one has to do something, but the question then is: what meaningful action can one take?

The fact that this question arises points to the failure of traditional university education. With its undue focus on teaching specific techniques, the more important second-order skill of learning to learn has fallen by the wayside.  In reality, though,  it is now easier than ever to learn new skills on ones own. When I was hired as a database architect in 2004, there were few quality resources available for free. Ten years later, I was able to start teaching myself machine learning using topnotch software, backed by countless quality tutorials in blog and video formats. However, I wasted a lot of time in getting started because it took me a while to get over my reluctance to explore without a guide. Cultivating the habit of learning on my own earlier would have made it a lot easier.

Back to the future of work

When industry complains about new graduates being ill-prepared for the workplace, educational institutions respond by updating curricula with more (New!! Advanced!!!) techniques. However, the complaints continue and  Bateson’s notion of second order learning tells us why:

• Firstly, problem solving is distinct from problem formulation; it is akin to the distinction between human and machine intelligence.
• Secondly, one does not know what skills one may need in the future, so instead of learning specific skills one has to learn how to learn

In my experience,  it is possible to teach these higher order skills to students in a classroom environment. However, it has to be done in a way that starts from where students are in terms of skills and dispositions and moves them gradually to less familiar situations. The approach is based on David Cavallo’s work on emergent design which I have often used in my  consulting work.  Two examples may help illustrate how this works in  the classroom:

• Many analytically-inclined people think sensemaking is a waste of time because they see it as “just talk”. So, when teaching sensemaking, I begin with quantitative techniques to deal with uncertainty, such as Monte Carlo simulation, and then gradually introduce examples of uncertainties that are hard if not impossible to quantify. This progression naturally leads on to problem situations in which they see the value of sensemaking.
• When teaching data science, it is difficult to comprehensively cover basic machine learning algorithms in a single semester. However, students are often reluctant to explore on their own because they tend to be daunted by the mathematical terminology and notation. To encourage exploration (i.e. learning to learn) we use a two-step approach: a) classes focus on intuitive explanations of algorithms and the commonalities between concepts used in different algorithms.  The classes are not lectures but interactive sessions involving lots of exercises and Q&A, b) the assignments go beyond what is covered in the classroom (but still well within reach of most students), this forces them to learn on their own. The approach works: just the other day, my wonderful co-teachers,  Alex and Chris, commented on the amazing learning journey of some of the students – so tentative and hesitant at first, but well on their way to becoming confident data professionals.

In the end, though, whether or not an individual learner learns depends on the individual. As Bateson once noted:

Perhaps the best documented generalization in the field of psychology is that, at any given moment, the behavioral characteristics of a mammal, and especially of [a human], depend upon the previous experience and behavior of that individual.

The choices we make when faced with change depend on our individual natures and experiences. Educators can’t do much about the former but they can facilitate more meaningful instances of the latter, even within the confines of the classroom.