Category: What Dimitri Learned

Because of the way we humans are wired, the expectation gradient is not a neutral measurement. For some reason, when we perceive a tiger eyeing us voraciously, our bodies immediately start pumping out adrenaline and otherwise prepare us to scale that gradient wall. In many ways, we literally transform into a different being. A thoughtful and kind individual is replaced by the instrument of survival driven by animal-like instincts.

But… who is doing the replacement? (Are you ready for the big reveal?) It would appear that we have another OODA loop, operating inside of us. Our body is running its own game, regardless of ours and with or without our awareness of it. Its intention is focused squarely on meeting demands of the expectation gradient.

This inner OODA loop is fairly primitive. It knows nothing about our aspirations. It cares very little about the intentions we form and write down in bold letters in decks and strategy 5-pagers. All it does is watch the gradient, trying to discern the gap between our current energy output, what the gradient says it should be, and try to change it as expediently as possible. Somewhere a long time ago, the evolutionary processes took us toward the setup where our unconscious mind is constantly and repeatedly asking this question: “How does the expectation gradient slope look right now and how much of my total energy do I need to mobilize to scale it?” 

For what it’s worth, such a two-loop setup is not uncommon. For example — you probably guessed where I am going — rendering graphics is a fairly computationally expensive process, and as such, makes processors heat up. To avoid overheating, most modern microprocessors have a tiny system called “thermal control” that’s built into most modern microprocessors… and it cycles its own OODA loop!

The thermal control loop is ignorant of rendering. It simply checks the processor’s temperature, and if the temp is above a certain value, takes an action of slowing down the processor’s clock. As a result, the rendering pipeline suddenly moves a lot slower and can no longer fit into the frame budget, producing jank.

It seems like a good thing, but more often than not, the result is deeply unsatisfying. The two loops are playing two different games, and step on each other’s toes, forming the familiar sawtooth pattern of jank. In consumer hands, this device seems downright menacingly janky. The brief periods of responsiveness feel like a taunt, like the device is actually messing with us. Back in the Chrome team, we’ve spent a bunch of time testing the performance of mobile phones, and many of those phones suffered this malady. As one of my colleagues quipped: “This is an excellent phone … as long as it’s sold with an ice pack.”

Similarly, our inner OODA loop is doing its thing, and the model of its environment is limited to the expectation gradient it periodically checks. Given that the expectation gradient is just a guess and often wrong, it’s no wonder that the inner, unconscious OODA loop ends up fighting with the conscious OODA loop we’re running, producing remarkable levels of macro jank.

From the perspective of the conscious OODA loop, this feels like a rug being periodically pulled from under us. I wanted to lose a few pounds… So what am I doing eating a Snickers bar in the pantry? I decided to work heads-down on a proposal today … So why am I watching random YouTube videos? We wanted our team to innovate daringly…  So what are we doing arguing about the names of the fields in our data structures? Ooh, a new Matrix movie preview… Stop it!

We might believe that we understand our intentions. We might even believe that we have a clear-eyed view of our “what should be.” Unfortunately, our simple-minded, yet highly effective, honed by eons of evolution inner OODA loop also has intentions. And these intentions, whether we want them or not, are woven deeply into the story of our actual “what should be.”

The conflation of “what is” and “what should be” is not the only way in which our intentions impact our prediction error rate. Another source is the intentions that we’re unaware of. To better understand what happens, we are going on another side adventure. And yes, we might even get to cast trigonometry spells again. But first, let’s talk about expectation gradients.

If we view prediction error rate as a measure of the accuracy of our predictions after the fact, expectation gradient is our forecasting metric. An easy way to grok it is to visualize ourselves standing on a trail and looking ahead, trying to guess the gradient of the incline. Is there a hill up ahead, or is it nice and flat? Or perhaps a wall that we can’t scale? The gradient of the path ahead foretells us of the effort we’ll need to put into moving forward.

In a similar vein, the expectation gradient reflects our sense of the difference between our models of “what is” and “what should be.” It is our estimate of the steering effort: how much energy we will need to invest to turn “what is” into “what should be.” A gentle slope of the gradient reflects low estimated effort, and as the estimate grows, the slope becomes steeper. If I find myself in a forest, facing a hungry tiger, I am experiencing a very steep gradient. Sitting in a comfortable chair while sipping eggnog (it is that time of the season!) contentedly and writing, however — that’s the definition of a gentle gradient slope for me.

With our trig hat on, we can picture the expectation gradient as the angle of a triangle. The adjacent side is the distance between “what is” and “what should be” (or a fraction thereof), and the opposite side is the measure of the required energy that we need to muster to steer the environment from “what is” to “what should be.”

The opposite-adjacent relationship to the angle is a tangent. When we deal with tangents, we face impossibilities. There is an asymptote, built into that little arrangement. The wavy tangent line starts slow, but then zooms into the sky, never ever quite fulfilling the promise of meeting required output.

I quite like this framing, because it feels pretty intuitive. The curve practically begs to be broken down into three distinct sections: the section before the kink where we’re reasonably certain that we can achieve our goal, the middle section where we are are uncertain of the outcome, and the asymptote – the section in which we’re pretty certain that our goal is unachievable.

Looking at “dancing with delusion” from the previous piece through the lens of expectation gradient, it’s all about convincing the team that the road ahead is mostly out of the third section, stretching the “uncertain” a bit longer.

The ability to make predictions is an astounding quality of retained-mode systems. Unlike the imaginary immediate-mode beings, we can start seeing what might be. And we humans are blessed/cursed with the ability to go one step beyond that: we can imagine alternatives. We create multitudes of “what might bes”. This is where the true significance of the Decide step becomes evident — we need to choose among the many things that “might be” to pick the Action that will do … what?

Turns out, we have preferences. We want some alternatives more than others. In our minds, the possible futures aren’t equal. A way to think of it is that we have a preference toward a certain state of the environment that we are imagining, as compared to our perceived current state. Instead of having just one model of the environment, we carry two: “what is” and “what should be.” The sum of our preferences manifests as intention, or our desire to move “what is” toward “what should be.”  

Every team is born and continues to exist around some objective. The objective itself may change over time, but it is the presence of this objective that holds the team together. This team, with some understanding of the environment —  “what is” — sets out on a journey of applying intention, to influence the environment toward some state, or “what should be.” From this perspective, the OODA loop is about steering — shifting the environment into some desired state.

Whoa, this is kind of a big insight, isn’t it? We’ve been walking around the OODA loop for a bit now, and — boom! — here we arrive at this moment. What is the point of the OODA loop if not saying: “Hey environment! I have some ideas about you. Let’s dance.” Within an OODA loop, intentionality is always present. There is no need for the OODA loop without it.

This whole steering business doesn’t come without downsides. Most individuals and certainly many teams don’t have the “what is” and “what should be” models clearly separated. One of my friends has a habit of pointing out the distinction, occasionally blowing people’s minds. This lack of separation commonly leads to the “what should be” model influencing the Observe and Orient steps. When Observing, we tend to filter out things that shouldn’t be there, and when Orienting, fit things that should be.

Guess what that does to our prediction error rate? That’s right. The more we the “what should be” model bleeds into our “what is” model, the higher is the error rate.

We live in a world where it’s hard not to notice the instances of conflation of “what is” and “what should be.” From conspiracy theories to magnetic personalities creating “reality distortion fields” around them, to filter bubbles, we are surrounded by them. We even have a term to describe some of these instances: cognitive biases. 

At the same time, this conflation can be productive. A team believing that they can ship a product may indeed ship a product, despite the overwhelming odds. Had they not mixed their “what is” and “what should be,” they would have seen right through their silly naivete at the start and folded early. Lucky for them, the environments are steerable – they can shift to “what should be” under certain conditions. The trick of any team lead is to recognize and hold the delicate balance between productive and unproductive blending of “what is” and “what should be.” Another friend of mine calls this balancing “dancing with delusion,” and I love how well it captures the nature of the process.

As it usually happens, we find ourselves in a conundrum. When managing jank, do we focus on the accuracy of our predictions or do we try to stay on pace with the clock? There does not seem to be a good answer — and trust me, “both” rarely feels helpful in the middle of the OODA cycle. It’s an iron triangle of seemingly impossible constraints. Given our current capacity as constant, we have to pick one of the two others: time or accuracy. 

Each presents two different configurations for the OODA loop: I’ll call them the OOda loop and the ooDA loop (note the capitalization).

Leaning toward the OOda loop, we spent most of our budget trying to perfect the model, favoring the Observe-Orient steps. We try to “consider all possibilities” and “look at the whole picture” when leaning toward this side of the spectrum. We hesitate to engage, hoping that the nature of the environment will reveal itself to us if we just keep our eyes peeled.

In the extreme, this configuration turns into the OO loop. We are subject to our “flight” instinct. We zoom out as wide as possible, trying to find ways out of the situation we’re currently in, gripped by the anxiety that comes with trying to consume the entirety of the environment. Everyone and everything is a potential threat, and every part of the environment hides nasty surprises. Every possible action looks like a wrong move. There is no escape.

This configuration produces jank that is immediately visible and seen, rarely a micro jank. Skipping a move is a big deal — and also a form of action. To collect more information about the environment for each iteration of the cycle, we need to act. Missing our opportunity to do so reduces effectiveness of our Observe-Orient steps. Despite our best and widest stares at the world, we are passive participants and our learning is limited to what is seen. The “analysis paralysis” is a common description of a team that is veering too hard onto this side.

In the ooDA loop, we forget — or willfully ignore — that the model might not be accurate. We concentrate our energy on the Decide-Act part of the process. If someone is calling for “bias toward action,” they are probably looking to move closer to this configuration. We lose sight of our model being just a fanciful depiction of the environment. It feels like “we’ve got it,” we finally “figured it out,” and now it’s time to seize the moment. All we have to do is “create order from chaos.”  

At the very extreme is the DA loop, when we’re driven entirely by our “fight” instinct. Here, our vision tunnels, and we only see simplified caricatures of the environment. A driver who just cut us off in traffic is a “stupid moron.” A colleague who said something we don’t agree with in a key meeting — a “backstabber.”

The ooDA configuration feels good at first. Asserting that the environment is “solved,” we gain a sense of certainty and confidence. Unfortunately, our prediction error rate tends to compound, because the model is being neglected — with each new cycle, we plow farther and farther away from reality. This compounding results in exponential growth in jank. We already know how this ends. From inside the organization, DA feels like one fire after another, sudden and unexpected. When teams are caught in constant fire-fighting and struggling to get out of one mess, then falling straight into another — chances are, they are favoring the ooDA loop’s end of the spectrum.

Neither of these extremes is a pleasant place to be, so organizations rarely spend time sitting in any of them. Instead, they lurch from one end to the other. The analysis paralysis gives way to “time for decisive action,” which is followed by “need to regroup and reassess” and so on. And in the process, teams pipe out jank like the smokestacks of the industrial revolution.

Individually, we all have our go-to OODA configuration as well. It is helpful to know our biases. For example, my first instinct is to shift to OOda, often in unproductive ways. Some folks I know prefer the more Leroy Jenkins style of ooDA, and recognizing how we might react in various situations helps us collaborate and reduce the collective lurching from one extreme to another. 

It seems that the retained mode is our way to compensate for the limited capacity to receive and process information about the environment. The implicit hypothesis behind the retained-mode setups is that we can make predictions based on the model we’ve constructed so far. As we Decide-Act, most of these will pan out, but some will generate prediction errors: evidence of incongruence between the model and the environment. We can then treat these errors as fodder to chew on in the Observe-Orient steps in our OODA cycle. Our rate of prediction errors for each cycle tells us how well we’re playing this whole OODA game.

Let’s see if we can add the concept of prediction errors to our framework. One way to visualize the idea of the model that is representative of the environment is to play on the idea of detaching from reality. You know, when we daydream about things at the stove, forget to turn down the heat, and burn our green beans (not that it ever happened to me). At that moment, our framework’s timelines come askew, with the environment’s timeline proceeding in one direction, and our model’s going in a slightly different one, at an angle.

Now, let’s say that the angle is informed by the amount of the prediction error generated during this OODA cycle. Allow me to channel my inner highschooler and do some arcane trigonometry: a triangle formed by the environment’s direction, and the model’s direction, and the adjacent-hypotenuse angle being the prediction error rate (kudos to my son for helping me remember all this nonsense).

There’s something very important about this relationship. With the environment clock continuing to tick at the constant rate, higher prediction errors will introduce a time dilation effect within the model: the clock will appear to be speeding up, leaving less space for the OODA loop to cycle! And what does that likely mean for us? Yup — more jank.

I will now take a tiny leap of faith here and correlate prediction errors and jank. Here it is: the higher our prediction error rate, the more incidents of jank we will experience. It seems that if we have a really awesome model that generates absolutely no prediction errors, we’ll have no jank. We’ll be like that youthful Keanu at the end of the Matrix, folding one of our hands behind our back, suddenly bored with the pesky Agent Smith. Conversely, if our model generates only prediction errors, it’s going to be all jank, all the time. We’ll feel like the Agents Smith in that same scene.

So it is likely that anytime we’re experiencing jank, we might be experiencing a troubling prediction error rate. Micro jank will come from the relatively small rate, and macro jank — from when the angle approaches 90 degrees (π/2 for you trig snobs) and the model clock is spinning like a top.

In either situation, especially when we feel like we have no time to react, it might be a good idea to reflect on how well we understand our environment — and most importantly, whether we’re aware that we only operate on the model of it. 

One of the most common mistakes organizations make is confusing high rates of prediction error in their models for the environment raging against them. If you ever had a fight with a loved one, and was humbled by recognizing how your assumptions took you there, that must resonate. With all the jank we produce and we’re surrounded by daily, and the enormous piles of prediction error rate this must represent, do you ever wonder how much slower the environment’s actual clock is compared to the one we perceive? And the untapped potential that the difference between them represents?

It will probably not come as a surprise to you that we humans are a retained-mode bunch. It’s cool to imagine ourselves as the immediate-mode beings: everything in the world around us would be brand new! For every cycle of our OODA loop, nothing is retained. Talk about living in the present.

Alas, — or fortunately, it’s hard to tell — we aren’t like that. It would totally suck if for every situation, we would need to relearn everything from scratch. We can only learn a tiny bit from each iteration of the OODA loop. Our strength, individual and collective, is in harnessing the retained mode. For example, when we look around the room, we can only see what’s in front of us. Yet we retain details of the room that aren’t in our direct eyesight, and can reason about them. We can reach for a glass of water without looking at it. This is our model being put to work. Every cycle makes the model a bit richer and more nuanced, helping us not just visualize things that we’re not seeing directly, but also make predictions about what happens to them in the immediate future.

When I first learned about the OODA loop, I naively presumed that all steps in the process operate directly on the environment. I observe the environment, I orient within it, I decide on what to do, and then I act on it. It wasn’t until later, after I learned about the concept of constructed reality, that a different understanding of the OODA process had emerged.

Aside from the first step, the OODA loop operates on the model of the environment, rather than directly on it.  This can be amazing, allowing us to connect our hockey stick with the puck for that awesome from-behind pass that sets the stands afire. It can also be a lot less awesome, because our models aren’t always representative of the environment. I reach for a glass — and accidentally poke it with my thumb, spilling the water. The model lied. 

Put differently, most steps in OODA occur in a mirror world of the environment that we created in our minds. If the mirror is clear, our actions proceed as intended. If it’s one of those funhouse mirrors, your guess is as good as mine. Our models are the sources of both our clairvoyance and our blindness.

Whether we want it or not, the OODA loop serves two interrelated purposes: one is to produce an action between the two ticks of the environment’s clock. The other is to update the model of our environment and keep it accurate. How well we manage to perform both tasks reflects in how we produce jank.

At the core of the OODA loop is the concept of a model. To create space for exploring it in depth, we’ll make a tiny little digression back into — you guessed it! — graphics rendering technology.

With my apologies to my colleagues — who will undoubtedly make fun of me for such an incredibly simplified story — everything you see on digital screens comes from one of the two modes of rendering: the immediate or the retained modes.

The immediate mode is the least complicated of the two. In this mode, the entirety of the screen is rendered from scratch every time. Every animation frame (remember those from the jank chapter?) is produced anew. Every pixel of output is brand new for each frame.

You might say: yeah, that seems okay — what other way could there be? Turns out, the immediate mode can be fairly expensive. “Every pixel” ends up being a lot of pixels and it’s hard to keep track of them, yet alone orchestrate them into user interfaces. Besides, many pixels on the screen stay the same from frame to frame. So clever engineers came up with a different mode.

In retained mode, there exists a separate model of what should be presented on screen. This model is usually an abstraction (a data structure as engineers might call it) that’s easy to examine and tweak and it is retained over multiple frames (hence the “retained” in the name). Such setup allows for partial changes: find and update only the parts of the model that need to change and leave the rest the same. So, when we want a button to turn a different color, the only part that has to be changed is the one representing the button’s color.

Both modes have their advantages and disadvantages. The immediate mode tends to need more effort and capacity to pay attention to the deluge of pixels, but it also offers a fairly predictable time-to-next-frame: if I can handle all these pixels for this frame, I can do so for the next frame. The retained mode can offer phenomenal benefits in saving the effort and do wonders when we have limited capacity. It also yields a “bursty” pattern of activity: for some frames, there’s no work to be done, while for others, the whole model needs to be rejiggered, causing us to blow the frame budget and generate jank.

This trade-off between unpredictable burstiness and potential savings of effort is at the crux of most modern UI framework development. The key ingredient in this challenge is designing how the model is represented. How do elements of the screen relate to each other? What are the possible changes? How to make them inexpensive? How to remain flexible when new kinds of changes emerge?

The story of Document Object Model (DOM) can serve as a dramatic illustration. Born as a way to represent documents at the early beginning of Web, DOM has a strong bias toward the then-common metaphor of print pages: it’s a hierarchy of elements, starting with the title, body, headings, etc. As computing moved on from pages towards more interactive, fluid experiences, this bias became one of the greatest limiting factors in the evolution of the Web. Millennia — hell, probably eons — of collective brain-racking had been invested into overcoming these biases, with mixed results. Despite all the earnest effort, jank is ever-present in the Web. Unyieldingly, the original design of the model keeps bending the arc of the story toward the 1990s, generating phenomenal friction in the process. 

In a weird poetic way, the story of DOM feels like the story of humanity: the struggle to overcome the limitations imposed by well-settled truths that are no longer relevant.

If our team’s OODA loop runs just a tiny bit slower than the clock of the environment, we will generate a flurry of micro-jank — many incidents that are so tiny, we can barely notice them. Unlike with machines, our collective resilience will helpfully wallpaper over these thousand cuts. However, as we’ve learned before, an incident of jank creates a deficit for the next cycle. It is fairly easy to see that this deficit continues to accrue over time. So the micro-jank grows into larger problems over time.

This larger problem usually manifests as macro-jank: a big reset that is clearly felt by everyone in the organization. The whole team seizes up and briefly stops listening to the environment’s clock, focusing inwardly to sort out their own mess.

In my experience, this phenomenon has an easily recognizable marker. A team that accrues OODA deficit tends to fall into this gait of periodically changing things around to see if their troubles will go away. However, because the source of the deficit remains unexplored, the rearranging of furniture rarely results in lasting change. Be it a dramatic shift in priorities, changing of leadership, or a reorg — it’s at best a temporary fix, quickly leading back to deficit accrual.

One of my go-to examples of this sawtooth pattern is “leads reset.” As the team forms, a small group of leads is organized. At first, these leads operate as an effective unit, providing valuable direction and insights on priorities to the rest of the team. However, as the time goes by, leads discover gaps in their knowledge, and pull in more people onto the leads group. Sometimes this happens as a result of a team growing, but often, the breadth of the challenge is such that a small group of people simply can’t grasp it fully. Plus, it feels important to be in the leads group. After a little while, the group of leads becomes large and unwieldy. Effective conversations yield to bickering and eye-rolling. Leads themselves become disheartened, which percolates throughout the team. So what happens next? As you’d probably guessed, a new, smaller group of leads is formed — until the next reset.

Having been part of these groups and an organizer of them, it always struck me as weird: why is it that we keep trying this same method to organize a leadership structure, over and over again? When a question like this pops up, it’s a good sign that the OODA deficit is being accrued.

Can macro-jank happen spontaneously, without first accruing micro-jank? It seems possible. Like, let’s imagine a severe and rapid environment change… oh wait, we don’t have to. It’s right outside. The COVID-19 pandemic will likely be a subject of many studies as a dramatic disruption of our environment. But was it truly an unexpected event or rather an outcome of micro-jank accumulating over a long period of time? How might we reason about that? To get there, we need to take a closer look at the nature of the OODA loop.

Putting jank and the OODA loop next to each other, it’s hard not to see the similarity. Both have two timelines racing against each other. Both describe one timeline trying to go a bit faster than the other — and there’s one timeline that consists of a repeating sequence of steps.

The question that got me excited was: “What does jank look like for the OODA loop?” To answer it, I did some light reframing to express the OODA loop as the timeline view that we’ve learned from rendering animations.

In this timeline view, the environment’s OODA loop becomes the ticking clock. Within each tick of this clock, we fit the familiar pipeline-like process: observe, orient, decide, act.

This setup is not exactly the same as having two independent nested loops, interacting with each other. However, since we’re most interested in situations where the nested loops are closely matched, this simplification works well enough. Here, the environment sets the pace and we try to match it. With each tick of the clock, the next round begins.

In this framing, the OODA jank is the situation when the cycle of the inner OODA loop is taking longer than the one of the environment. 

Since I’ve just concocted a different way to look at the OODA loop, I might as well add another twist. Unlike in Boyd’s original military context, OODA jank is not lethal for most organizations. It is something that happens commonly, perhaps many times over.

Team jank is not great news, but most of the time it’s not existential, either. Deadlines get missed, but things still get delivered. People are late to meetings, but they do show up. Product launches get delayed, but most often, still happen at a later date. Human systems aren’t mechanical. They tend to be more resilient to jank. When rendering Web pages, a late-to-render animation frame is completely dropped. In organizations, being late just means reduced effectiveness, a miscalculation and wasted energy.

As a result, organizations typically feel jank not as one specific incident, but rather as cumulative effects of multiple instances. Perhaps we can use this newly-minted framework to dig into these effects?

I first learned about jank when I joined the Chrome team. It’s a weird slang word with multiple meanings,  so I am going to use a narrower definition, custom-crafted just for this narrative. To get there, I will make a brief detour into the land of rendering Web pages. Hold on to your hats.

Suppose you are visiting a site. I’ll be the browser in this story. You just clicked on a button, and I need to play out a lively animation as a result. Like humans, browsers are mesmerizingly complicated, but at a very high level, the animation is a sequence of frames — pictures of the intermediate states between its beginning and end. Each frame is rendered — that is, created on demand in a very brief moment of time. For example, to play out an animation at a common-for-computers rate of 60 Hz (that’s 60 frames per second), I have just under 17 milliseconds to render each frame.

Rendering itself is a multi-step process, usually called a pipeline (does this start to remind you of something?) To produce a frame, I must go through each step in the rendering pipeline. Think of it as a clock that ticks every 16.667 milliseconds. If I was able to fit all the steps between the two ticks, I have a frame of animation that I can show to you. Yay!

However, if going through the rendering pipeline takes longer than that, the next tick will arrive before I have the frame ready. Bad news. Despite all the work that I’d done, you won’t see this frame. It’s dropped. Worse news: because I had to finish all the steps (those are the pipeline rules), I accumulated a deficit — my work on the frame that follows begins with the negative time balance. For example, instead of 16.667 milliseconds, I might only have 12. What’s the likelihood that this frame will get dropped as well? Pretty high.

As a user, you will see this phenomenon as “jank”: instead of a smooth animation, it’ll look like a stuttering janky mess. Put very dryly, jank is the observed effects of a regularly scheduled pipeline-like process not fitting into its allotted time budget.

Wow. That is very dry. Let’s see if we can make it a bit more useful by applying what we learned here to the OODA loop. Let’s unroll the OODA loop.