Education's Bitter Pill - I

In a very short piece written in 2019 called "The Bitter Lesson", Rich Sutton of the University of Texas at Austin made the observation that human intuitions about how best to program computers to solve problems and play games have been shown to be flawed. In particular, the approach that believes that what we have to do is reproduce in machines - specifically in different versions of AI embodied in Large Language Models (LLMs) - the same thinking-processes that we believe human beings to engage in.

This is what he writes:

We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.

Now consider this point in relation to education where almost all our efforts are dedicated to building knowledge into children/pupils/students. External agencies - usually governments through the controlling mechanisms of examination régimes - decide what needs to be known; controlling agencies such as educational institutions and examination boards assess whether what needs to be known has been learned in ways they deem appropriate; life-defining agencies such as employers and societies then decide whether the learning has been done to a sufficient standard and in sufficient quantity and quality to merit reward in terms both of role and remuneration.

But suppose that building knowledge into students is as futile as building it into artificial intelligence engines, that whatever gains are made for students, institutions, societies and nation-states are only beneficial in the short term, but in the longer term only inhibit further progress by incarcerating every successive generation in the standards and ambitions of the past. By definition, what can be taught and learned on such a scheme is what is already known. Those who succeed best turn out not to be those who have most wholeheartedly acquiesced in this and 'successfully' navigated its obstacles, but those who have rejected it and instead followed their own interests, abilities and ambitions.

One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.

Call this the 'technological paradox', that in an age where technical knowledge and engineering skill are hugely value and highly rewarded there is an almost irresistible temptation to seek the kind of know-how that furnishes immediate returns and rewards, but that in the longer term this inevitably proves counter-productive because the methods that lead to such short-term success do not scale and do not afford the innovative, creative capacities that solve larger, harder problems. Going for short-term gains based on necessary technical know-how almost certainly limits our capacity to make long-term improvements because we have converged on something immediate rather than cultivated the skills that produce longer-term insights and rewards.

The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity. Essential to these methods is that they can find good approximations, but the search for them should be by our methods, not by us. We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.

The emphasis here is mine: "we should build in only the meta-methods that can find and capture this arbitrary complexity". In other words, as educators we need to swallow the bitter pill that by concentrating on content at the expense of 'meta-methods', on rote learning and copious factual knowledge easily listed and just as easily tested, we have 'sold the pass', going for the obvious at the expense of the important and valuable. We have spent millennia, centuries, decades and years - choose your own time-span - attending to precisely what does not matter.

But let us go back to an earlier phrase Sutton uses and see its real significance, that "We have to learn the bitter lesson that building in how we think we think does not work in the long run". The point here is the "think we think" because the way we think we think is hugely dominated by our predilection for attributing powers of the mind to conscious processing. We like to think - our default assumption is - that thinking is deliberate and conscious when it is neither. The best we can do is to use the very limited bandwidth of our consciousness to tell our non-conscious brains to attend to some area or specific problem. Thereafter we are compelled to "wait and see". Sometimes the result of our non-conscious thinking comes quickly in a conscious thought; sometimes it takes a while; sometimes it doesn't come at all. But what is certain is that between the "let me think about this" and the "how about this answer" our consciousness plays almost no part - and usually absolutely no part - in our "thinking". Everything is non-conscious.

Of necessity, therefore, one of the most powerful lessons we learn and skills we develop when it comes to problem-solving is the ability to let go, to leave our non-conscious brains to their invisible and unintelligible deliberations, and acquire the patience to wait for the result. The folly of strictly time-limited examinations will be apparent in this inversion of our usual practice.

So we have not only tried and failed to program computers to solve problems and play games using brute force and detailed databases of knowledge: we have failed to educate ourselves to do so for just the same reasons. We have fallen for the trap encapsulated in the aphorism because we cannot measure what we value we value what we can measure, and what we can measure is what we can quantify, list, test, rank and use to discriminate between human beings on grounds that are, by the same argument, mistaken, unjust, inefficient, unsuccessful and wasteful of human lives.

We remark as an aside that if we want to find an explanation for all this mistakenness we need look little further than another human propensity to believe that people are intrinsically better than one another, a belief itself one of the worst consequences of the doctrine of souls, hierarchically-ordered and possessed of 'gifts and talents' that differentiate them from birth or conception. The education system is supposed to identify, assess and rank without fear or favour, but in fact the way it is organised serves only to confirm this intrinsic, predetermined hierarchy, to reinforce the ability of those who can do what it thinks needs, and so governments and societies think need doing so that they produce new generations of hierarchically-organised social beings capable of doing exactly the same. Hierarchy so conceived is a self-reinforcing, self-sustaining delusion: we praise and reward those who can do the things that those who have decided what is worthy of praise and reward can do. We grant the greatest accolades not to those who can identify and solve important new problems, but to those best able to regurgitate extant solutions to old ones.

It is a fair counter-argument to say that until recently we had little option but to store information in human brains because the alternatives - mostly books and manuscripts - were impossibly difficult to access in volume and at speed. But that ceased to be true during my lifetime, and I see few if any signs that we are making any attempt to migrate to a better education system because most of the people who run education are vested in the status quo and most governments too entrenched and self-absorbed to challenge prevailing social practices.

But much more needs to be said ...