On the Psychology of Architecture and the Architecture of Psychology

If you’re an architect — in my world is that a ‘digital architect’ working on complex Business-IT landscapes — the concept of convictions is key. It is of course true that the added value of an enterprise architect is based on the capability to form valuable opinions — where valuable has many forms, but mostly they are surrounded around the idea that the opinions help to improve the organisation. So, a good architect must be able to form sound convictions and judgements — there, the job of enterprise/digital architect becomes intertwined with technology. But a good architect must also be able to convince others of those sound convictions and judgements. It is there that the job becomes intertwined with the psychology of humans, both the architects themselves and the people they interact with.

We might put the architects in the following scheme (of course, the table represents a more fluid reality) with respect to the influence of sound technological judgement on the one hand and being able to convince on the other:

THE ADVISOR	Poor Judgement	Sound Judgement
Not Convincing	Mostly Harmless, Potentially Harmful	Potentially Valuable
Convincing	(Very) Harmful	(Very) Valuable

Many architects are wrestling with the key question of being heard. They are — as we all mostly are — convinced they have that sound judgement. Others may differ on that conviction of course, but you don’t become an advisor unless you have a decently sized ego (cough). But how do architects actually get their ideas accepted and acted on? Often, architecture conferences have discussions on ‘being heard’, ‘getting the ear of executives’, and so forth. This was also the case on the conferences I was at recently. Generally, these talks and discussions focus on the architect themselves: they have valuable input, but what can they do to be heard, to be valuable, to be effective, to convince?

‘Convincing people’, however, in reality depends a lot on the receiving end: the people who the advisor needs to convince. And fairly recent psychological and neurological research has given us quite a bit of insight in that side of the equation, and it turns out our convictions about human intelligence have been mostly wrong for centuries.

Because for centuries, millennia even, we have tended to believe that humans create convictions (are convinced) on the basis of observations, ‘facts’ and logical arguments. But it has turned out that observations/facts and logic only have a very limited effect on our convictions. This most clearly becomes visible these days by the fact that conspiracy theories are rife, and sometimes batshit crazy, while people who are convinced of these conspiracy theories are immune to facts and logical arguments. And this immunity is not a matter of the level of analytic and logical talent. No, there are highly educated analytical and logical people who are deeply convinced of crazy theories, from a flat earth to human society being run by human-flesh eating lizards in human form. Others may call these people either ‘crazy’ or ‘dumb’, but I don’t. For me, they are simply very human and not that different from the rest. And while ‘rational’ ‘intelligence’ does provide some protection against crazy convictions, it is by no means effective. This is because — as psychological research has shown over the last decades — we humans tend to simply dismiss logic and observations that are in conflict with the convictions and assumptions that we already have, and the stronger our convictions are, the less effective logic and observations are able to weaken them (and the more effective logic and observations are able to strengthen them).

Illusionists, for instance, make use of this property of human minds. You can make sure the attention/consciousness is strongly focused elsewhere, in which you get the ‘Invisible Gorilla’ effect. But even without that trick, we can still be pretty consciously blind. For instance, the illusionist may first in some way convince people that something is the case, then they present situations that are in contradiction with that conviction. These situations are (unconsciously) observed — again, as research has shown — but these observations are then literally unable to reach our consciousness. Really: some things that are right before our eyes will be unable to reach our consciousness. If you want to read in depth about this, I suggest starting with Stanislas Dehaene’s Consciousness and the Brain, and The Invisible Gorilla from Chabris and Simons. There is a lot going on in parallel in our brains, but our consciousness/attention is ‘single-threaded’.

If the brain gets outside information (observations, information in general) that is ambiguous, then multiple different interpretations are available in the brain at the same time, but only one makes it into our consciousness at any one particular moment. And it is our existing convictions that have the strongest influence in what interpretation/observation reaches our conscious attention. That way, at least, existing convictions are easily reinforced, but not so easily changed. The brain even seems to ‘crave’ reinforcement of its convictions (we tend to ‘enjoy’, get pleasure from, confirmation that we are ‘right’).

The fact that observations and logic have little impact on our convictions — and that therefore our convictions are stable — merits attention. The question is: why? Well, it seems to me because there are key advantages of having stable convictions. These are:

• Stability of convictions makes us fast and efficient;
• Stability of convictions makes cooperation in (even very large) groups possible.

With respect to the former: it is simply not doable to analyse everything all the time*). You can’t go back to all sorts of ‘first principles’ when you have to scour the environment for predators or food, for instance.

With respect to the latter: stable convictions and assumptions also make ‘groups of us’ much more efficient. If the people in your environment would change their convictions and assumptions every day, you would have to spend a lot of time and energy to find out what their conviction today is, as you cannot assume that it is the same as yesterday. Here too: it is very inefficient and thus ineffective to doubt everyone all the time.

In other words, conviction-stability enables a higher operational speed of individuals and less friction (and thus more speed) of a group. Shared convictions enable efficient and thus effective groups, and these groups can even be be huge: we have nations of more than a billion people and these can act together only because they have shared convictions (and hence: attacking a nation by undermining shared convictions is a very effective form of warfare — digressing again…).

Which brings us back to that key question: how do we get and change our otherwise stable convictions? Because we do create and actually do change our convictions over time. Some of this is based on our ability to (slowly) create new patterns out of logic and observations. But the most effective way is to be exposed to either a lot of repetitions of something or something that comes from a source we trust. Trust is an assumption. I trust good science journalism because I assume the writers want to write trustworthy stories. Most powerfully influencing, though, is getting information from a direct interaction. Direct interactions tend to be implicitly trusted and thus information we receive from them is almost automatically accepted. It is part of what enables us to make effective (fast-acting) stable groups. Our (linguistic) intelligence is a perfect match for creating stable and effective groups of 150-200 (Dunbar’s Number) and this is visible even today as hunter gatherer societies have groups of maximally that size.

In other words: if you want to be a convincing advisor, psychological research tells us that your observations/facts and rational arguments may have relatively little effect. You need to make sure your relation with the person you need to convince is as good as possible and the message is heard as often as possible. Problem one: your message may be in conflict with already held convictions, this is especially true when there is a knowledge gap. Problem two: frappez toujours (keep pushing) might work against the relation as well — a real Catch-22. So it is above all important to try to get a kind of natural seat at the table where the key decisions are made. Be part of the group. But it remains often a chicken and egg situation.

Above, I wrote about the power of reinforcement. Have a look at the video below (it starts at 14:55 in, watch it until 16:06, so just over a minute)

I especially find the story about 70% of the persons in a trial getting convinced that they had participated in a crime that in reality never happened quite telling. Also interesting in that episode: memories are not like recordings, they are the ‘stories’ supporting our convictions. So, if our convictions change, we are psychologically in need of different memories. Which really happens.

Coming back to the role of advisor, it is clear we have to think about the psychology of people we want to convince, our ‘target audience’. And we have to take into account:

• Both their and our existing convictions are stronger than any argument or fact or observation the other can provide;
• Both their and our convictions may have small beginnings but become powerful on the basis of reinforcement. Both the volume/repetition and the closeness of reinforcement are key.

And in the opposite direction: if you, reader, read this and you are someone who is mostly an advise-taker (e.g. you’re a director or a company board member): does your chief legal counsel understand the ‘bits and bytes’ of the law? Does your chief IT advisor understand the bits and bytes of IT? Or is your trusted advisor above all trusted because they are close to you and spoken/seen often? Or trusted because they reinforce you with what you already were convinced of? Are contrarian views and observations part of your setup? Not? Well, that is only very human. Not necessarily very good, though.

These days, I tend to see this as the ‘architecture of human intelligence’:

A very small part of our intelligence is analysis and ratio. It is where our ‘attention/consciousness’ is too. This is where we find (most of?) what Kahneman in Thinking, Fast and Slow calls our ‘System2’: slow and deliberate thought.

Most of our intelligence, however, consists of patterns that we execute efficiently, automatically and quickly. Some of these are natural elements, which are fixed: e.g. a propensity to communicate and use tools**), to perform ‘mental travel’***) — memory, scenarios, fantasy — and all of it based on pattern creation and reinforcement. Some of these elements may even be genetic (like basic strategies such as wait-and-see versus go-for-it you can observe in small children), but most of it is probably learned. All of this is part of Kahneman’s ‘System 1’. We learn by employing our capability to employ logic and ratio and our copying-and-being-reinforced capability — and while we do a lot more copying than observing and analysing, culturally, we tend to believe that the reverse is true. I say ‘culturally’, but it might even be deeper than that: our conviction that our convictions are rational (and should thus be trusted) also results in more stability of the convictions. So, who knows, maybe our self-delusion is a survival trait.

Learning by reinforcement also includes learning by doing. Chess grand masters have very effective fast ‘patterns’ in their brains, and the difference between grand master and good amateurs is not their power of logic and ratio — calculating, thinking moves ahead — but their patterns that identify potential good moves before they start to calculate (research by Adriaan de Groot which showed that professional chess players thought roughly the same amount of moves ahead as good amateurs, Gary Kasparov once said that he normally calculates 3 to 5 moves ahead — except in special circumstances where many moves are forced — i.e. have no alternative), and these patterns come from playing a lot of games. You also have to maintain your patterns: it is ‘use it or lose it’.

A book I really liked — Predictably Irrational by Daniel Ariely — gives a nice example of the lack of power of the rational. In the book, Ariely compares two kinds of situations: one is that you have to walk 10 minutes to save $2 on a $5 investment and the other is to walk 10 minutes to save $2 on a $150 investment (not his exact example). It is the same effort ($2 for 10 minutes walk) but the first one we do and the second one we do not. Ariely shows this as obvious irrational behaviour, and tells us it is driven by relativity (saving $2 on $5 is a lot ‘more’ than saving $2 on $150). But following the model above I think there is another possible aspect in play. Chances are that the $5 investment is something we do often (e.g. luxury coffee) whereas the $150 investment (e.g. a suit) is something we do seldom. For something we do often like buying a coffee it pays to act on the saving, for the suit less so. The individual choice may not have been rational, but following the pattern is. This is supported by my observation that when we are in non-standard situations (say a vacation abroad) a pattern suddenly is no longer followed. E.g. on vacation, we will happily pay a lot for something we would never pay that amount for at home. It would be fun if some psychologist or behavioural economist would test this sometime.

Anyway, this brings me a to my ‘short definition of human intelligence’:

Human intelligence can be best described as ‘malleable instinct’

Me. Now.

Finally, I want to mention another book: David McRaney’s How Minds Change. In it you will find descriptions of techniques that have been developed over the last decade to have more than ‘hardly a chance’ to change people’s strong convictions. (Spoiler: most effective is to have people question themselves).

Next time if someone asks me what area is most important for a digital/enterprise architect, I’m probably going to answer: “Next to actually having a good technical insight? Psychology!”.

PS. In case you were wondering (and also in case you were not): our craving of and sensitivity to reinforcement is why (a) ‘social’ media companies can keep us hooked and make money of us, and (b) let them help fragment society into groups with separate convictions. The fact that social media is so good at influencing our convictions may also come from the fact that there is a lot of experiences that mimic ‘close relationships’. That influencer seems to be your ‘friend’ and in the comment section you will meet many more. A cocktail of the emulation of close relations coupled with positive reinforcement. All for profit. What could go wrong? But that is another story.

Appendix: IT versus humans

Alan Turing already observed that strictly speaking, digital machinery does not exist. It’s all analog. A transistor is a very analog thing. But we can create analog machines where we can ‘ignore’ everything except the stable states (that’s what a clock in a digital CPU is for). As Turing observed: “there are many kinds of machine which can profitably be thought of as being discrete-state machines”. Our brain can do the same, but we’re much less capable doing that than modern transistor-based digital IT is.

If our intelligence is not based on ‘logical analysis’ (this has been clear since the failure of the first wave of AI) but is (mostly) ‘malleable instinct’, then this is closely related to the fact that the brain is in fact ‘malleable hardware’ (‘wetware’). Changing patterns means adding and removing physical connections between neurons ****). To add to the definition above: Human intelligence can be best described as ‘malleable instinct’ implemented as ‘malleable hardware’.

In IT we do not have ‘malleable hardware’, but ‘malleable software’ running on ‘fixed’ hardware. Yes, we also develop new hardware, but that is a very slow (years for a change) and mostly incremental (faster versions of the same hardware logic) process. And when there is really new behaviour of hardware, migration of software is slow too. Hardware generally is a lot faster than software, which is why often the hardware changes are task-specific hardware-accelerations: tasks moving from slow software to fast hardware. Examples:

• In the 80’s we had hardware acceleration for calculating with (fake) reals (the ‘mathematical co-processor’). As of the 90’s a standard embedded element of any CPU (aside: very funny, there are modern processors that have no — expensive — division operator in hardware but that depend on trickery in the compiler, if you’re in to deep tech and calculation trickery: watch this YouTube video);
• In the 90’s we saw the rise of specialised graphical processors, the GPUs, which are now often used as parallel-computing workhorses for uses like building machine learning models like GPT-3 and proof-of-work based blockchain such as bitcoin-mining;
• In the 00’s there was the rise of hardware-acceleration for compression/decompression (video, e.g. H264, something that played an important indirect role in Apple’s iPhone success);
• And these days we have hardware-acceleration for encryption as part of our CPUs, which enables fast encrypted transport (https, VPN) almost without a performance penalty.

In the IT world we generally separate the ‘run’ from the ‘change’. The ‘change’ in most of IT — everything except hardware development — consists of programming. The result is employed in ‘operations’. Some of that development we have even started to call (machine) ‘learning’. Machine learning on digital IT is nothing but the ‘data driven programming of (hidden) algorithms’. The result of that machine learning (GPT-3, LaMDA, DALL•E, etc.) is a ‘rule based system in disguise’. Lots more rules than we can consciously create and maintain as humans, though, which fools us into thinking that someting fundamentally different is happening under the hood. Did I mention already how gullible we humans are? I think I did…

In one way, the human brain is indeed just like that digital IT: development in IT is like the learning by humans (and in extension by organisations). But there are two fundamental differences:

• As mentioned, the fact that our brains are based on malleable hardware, not on fixed hardware and malleable software. Our IT systems have a limited and fixed set of ‘hardware accelerators’ for only some specific algorithmic tasks (as mentioned above), our brains on the other hand are almost nothing but optimised hardware and almost no software. Our brain is one big generic hardware-accelerator, so to speak.
• But there is another more fundamental difference that I am convinced is essential to understand the limits of digital computing: the world of digital IT is purely logical, it is based on working with integers. An unimaginable amount of integers by now (expressed as the most fundamental integer type: the bit) and classical logical operations on these. That is what the IT-Revolution consists of. All of it. You may think your spreadsheet platform can do ‘real’ numbers, but that is an illusion. Real numbers are faked on digital computers by the clever manipulation of very large integers. Human brains, on the other hand, are not digital at all. Our abilities are based on reals, not integers. And you cannot express the true power of reals with integers. The impact of reals can be said to be 2 to the power of the impact of integers. That is why — as far as I’m concerned — current digital machine learning is a dead end on the road to true intelligence and why a 25 Watt human brain massively outperforms a 125,000 Watt digital brain (GPT-3) on most tasks in the real world. A single training session of the GPT-3 ML AI model already did cost about as much energy as a single human brain consumes in more than 100 years (and all of that to produce ‘brittle autocompletion on steroids’). Analog and quantum computing do work with operations on reals. Which is why researchers really want to get QM-Computing going so badly. Even Elon Musk sometimes admits (e.g. TED April 2022 Interview) that trying to create self-steering cars keeps on throwing up unexpected barriers (which I think it will keep doing until we no longer try to do it purely digital — not that unexpected if you ask me).

*) In fact, I have long suspected that the relation between certain forms of autism and being more analytical has to do with the fact that if you are analysing all the time, it is very hard to have stable automatic convictions that can survive the onslaught of confusing input, but I digress, as usual. (return to text)
**) And yes: Uncle Ludwig was right: language is a tool, and I’ve read suggestions by researchers that tool use and language might be pretty closely related — it might for instance be why we tend gesticulate when we speak — more digression… (return to text)
***) Watch that documentary. Memory is ‘fantasising/creating stories about the past’. And memory champions may become champions by creating ‘stories’. (return to text)
****) For long term-patterns. For short-time patterns we may quickly strengthen or weaken existing links as neurons may behave like a memristor (note: memristor is a concept that has seen various levels of fundamental critique but also interesting engineering). The brain really has an amazing architecture. (return to text)

This article is also part of the Understanding ChatGPT and Friends Collection.