Tools for Forgetting
I often find myself saying that it's impossible to fail, so long as you learn something in the process. The idea that acquisition of knowledge always results in a permanent step forward is inherently appealing to my brain. On the other hand, my life continues to slap me in the face with the value of unlearning. It seems that letting go of preconceived notions, to make space for the present moment, often results in radical breakthroughs in thinking.
Yet, restriction breeds creativity. Structures require scaffolding to take shape. It seems there is a delicate balance between learning and unlearning if we are seeking the best possible outcome. I suspect that this cycle of learning and unlearning is what makes generally intelligent agents (like, you know, humans) possible. Adapting to novel and changing environments is humbling, we must begin the learning process over again and our ego can get badly bruised in the process.
Adapting to a new environment requires us to relax our mental constraints. Only then, when we see clearly, can our internal systems settle into equilibria with a new set of constraints until, of course, the cycle repeats once again. For any system to evolve, some kind of shift in constraints has to occur in this iterative manner. This could be a change in the environment containing the system, or in the set of affordances the system can utilise to interact with the environment. Okay, so you probably get it by now, I'm arguing that this is some kind of fundamental principle that appears everywhere. This cycle of forward and backward, restriction and relaxation, tension and release is the nature of learning. Beyond the specific biological context, evolution and adaptation are simply other terms for learning in a changing environment.
This is all a roundabout way of saying that learning is not something that merely happens to us passively, nor is it strictly a property of biological brains. We can put in place and remove constraints intentionally to shape the development of any system, including our own bodies and minds but also our societies and machines. In fact, a buildup of constraints without ever relaxing or pruning them is what causes a system to stagnate and ultimately fail.
Anyone who's written a sizeable computer program knows that a steady accretion of rules is the path to ruin. We need to refine, remove, reorganise and refactor to shift our constraints and enable the system to evolve, without collapsing under its own weight. The same is true for any complex system of rules, perhaps this is most obvious in organisations:
Frustratingly, this principle means that at every point through the arc of time we will be putting in place constraints that will later hold us back. There is no direct route to a goal through a changing environment1.
Stability is a lie
In Tension and Freedom I argued that stability of any sort is always temporary, but I sense that most of us believe the world will stay roughly the way it is, forever. There is some reason to think this, we've created large pockets of relative stability in the otherwise writhing sea of chaos. I am lucky enough that the threat of starvation has been replaced, for me, with the mild inconvenience of walking to the shops. Often these pockets are stable enough to fool us into thinking change can be prevented indefinitely, that we can control our environment.
We tend to look back on the people of history as idiots. How could they not see the changes coming? Slavery is obviously wrong, colonialism is obviously wrong, fascism is obviously wrong etc. and, to be fair, many of these findings seem obvious to us today. Much more will be obvious tomorrow, and every day following. So, what will we be judged for? Farming and eating animals? Addiction to consumerism? Neurotic and dysfunctional political systems? Probably all of that and more, I imagine2.
The people of history, like us, thought they understood their world but we are mere links in the extremely long chain of life. As we store ever more information, probe deeper about the universe and march diligently towards creating digital demigods we must be careful not to become too certain of ourselves.
I often marvel at the complexities of the modern world, it's mind-blowing that I'm able to reach you (dear reader) from the comfort of my home. Humanity has achieved incredible things and we should be proud of both our ancestors and our peers today. We have learned so much as a species that it's become tempting to rest on our laurels.
Clinging tightly to our identities and our beliefs about what is right will hold us back from our potential, both as individuals and as a species. And, yet, this feels incomplete. Actively shedding all beliefs leads to literal anarchy3 and nihilism. How can we know the right amount of change? There is no stable answer here, there is only a balance to maintain.
We must meet reality at the pace it arrives, making an active effort to see it with fresh eyes in each moment. This isn't like the last time, it's always a little different. Learned shortcuts like legal precedent and peer-reviewed study are incredible tools that must also be regarded with great care. Any mental shortcut is both an asset and a liability, depending on the context.
Just as creators succumb to audience capture, so too can our technological systems, organisations and governments. Our media promotes a narrative that stirs up the people which create a new narrative to exploit, without anyone even noticing. We teeter on the edge of terrifying positive feedback loop, gradient-descent-ing our way into the abyss.
Tools for Forgetting
So, perhaps there is such a thing as too much optimisation4? Overfitting is a well known issue in data science and simulated annealing is a possible response5. Annealing, simulated or otherwise, is the process of injecting energy into6 a system to allow constraints to relax and reform into a more optimal configuration. The system partially “forgets” its original structure which enables further optimisation. This works especially well when done iteratively, which blacksmiths have known for a couple millennia. Applying sufficient heat to metal allows chemical bonds to break and, as it cools, the internal structure finds a new, stronger atomic configuration.
Our brains do something akin to this internally. The most significant moments in our lives are significant because of what we learn from them. As time passes we gradually forget the details and the emotional potency of past events and integrate them. However, it's certainly possible for our minds to be "stubborn" in the face of new information. Indeed, some kind of unwanted "stubbornness" seems to be responsible for many mood disorders, improper movement patterns and wrongheaded ideas.
Thankfully there are many tools we can use to deliberately trigger the annealing process in our minds: travel, contemplative practice, cognitive behavioural therapy, psychedelic drugs, loss, love and art are all stimuli that can lead us to “high energy” mental states. Unfortunately all of these come with a disclaimer, we have to be open to change in the first place. None of these tools can be used effectively without first adopting the mental posture of admitting we know almost nothing7.
If that wasn’t bad enough, there's yet another catch. These tools work at the individual level, not on organisations and societies. How can we prevent overfitting at the highest levels? I suspect we will need to engineer forgetfulness into our systems. There are some low hanging fruit here:
enacted laws could include clauses for re-evaluation and deprecation over time
manufactured products could require destruction and disposal plans at the time of construction
entertainment recommendation systems could allow users to partially or totally reset their training data and even manually edit it
In a broad sense the question for the coming decades on this planet seems to be “how can we engineer forgetfulness everywhere?” Returning to the body: all life is a balance between birth and death. Old cells die and new cells form in their place. Cells have to die properly, or else the entire organism can fail.
The purpose of thinking is to let the ideas die instead of us dying.
Alfred North Whitehead
Perhaps the most beneficial lesson we can leave for the people of the future is the importance of forgetting.
Until next time,
Ben ✌️
Related reading
I brought a lot of my ideas together in this one, it was a challenging one to finish. If you’re looking for further reading here are some sources that motivated this piece:
🧵 If you aren’t overcorrecting, you’re undercorrecting (Malcom Ocean)
🧵 Simulated Annealing Overview (metadiogenes)
🎥 Exploration-based Design (Casey Muratori)
📄 The Perils of Audience Capture (Gurwinder)
Stuff I’ve been thinking about
🎙 Jordan Peterson w/ Lex Fridman
🎙 Dr. Andrew Weil w/ Tim Ferris
🎙 Duncan Trussel w/ Lex Fridman
🎙 William MacAskill w/ Sam Harris
📄 Noosphere, a protocol for thought by Gordon Brander
📄 Moving beyond value from Mystical Silicon
🎶 As We Fall Into Deep Waters by Royal Coda
🎶 Feels Bad Man by Dance Gavin Dance
Just to be clear, I eat animals, buy things and participate in these dysfunctional systems. I will also be judged.
Which seems like an unstable state, someone ambitious seems to take over in these situations.
I think technically, rather than “too much” optimisation, this is optimisation with an incomplete model but show me a truly complete model
Caveat: I am not an ML engineer and have very little concrete knowledge of the space.
Or, more generally, increasing entropy. I usually try not to talk about entropy because I am bad at understanding it (for now).
This is known as “Beginner's Mind” in Zen