Pattern machines that we don’t understand

How do experts make decisions? One theory is that they generate a set of options, estimate the cost and benefits of each option, and then choose the optimal one. The psychology researcher Gary Klein developed a very different theory of expert decision-making, based on his studies of expert decision-making in domains such as firefighting, nuclear power plant operations, aviation, anesthesiology, nursing, and the military. Under Klein’s theory of naturalistic decision-making, experts use a pattern-matching approach to make decisions.

Even before Klein’s work, humans are already known to be quite good at pattern recognition. We’re so good at spotting faces that we have a tendency to see things as faces that aren’t actually faces, a phenomenon known as pareidolia.

As far as I’m aware, Klein used the humans-as-black-boxes research approach of observing and talking to the domain experts: while he was metaphorically trying to peer inside their heads, he wasn’t doing any direct measurement or modeling of their brains. But if you are inclined to take a neurophysiological view of human cognition, you can see how the architecture of the brain provides a mechanism for doing pattern recognition. We know that the brain is organized as an enormous network of neurons, which communicate with each other through electrical impulses.

The psychology researcher Frank Rosenblatt is generally credited with being the first researcher to do computer simulations of a model of neural networks, in order to study how the brain works. He called his model a perceptron. In his paper The Perceptron: a probabilistic model for information storage and organization in the brain, he noted pattern recognition as one of the capabilities of the perceptron.

While perceptrons may have started out as a model for psychology research, they became one of a competing set of strategies for building artificial intelligence systems. The perceptron approach to AI was dealt a significant blow by the AI researchers Marvin Minsky and Seymour Papert in 1969 with the publication of their book Perceptrons. Minsky and Papert demonstrated that there were certain cognitive tasks that perceptrons were not capable of performing.

However, Minsky and Papert’s critique applied to only single-layer perceptron networks. It turns out that if you create a network out of multiple layers, and you add non-linear processing elements to the layers, then these limits to the capabilities of a perceptron no longer apply. When I took a graduate-level artificial neural networks course back in the mid 2000s, the networks we worked with had on the order of three layers. Modern LLMs have a lot more layers than that: the deep in deep learning refers to the large number of layers. For example, the largest GPT-3 model (from OpenAI) has 96 layers, the larger DeepSeek-LLM model (from DeepSeek) has 95 layers, and the largest Llama 3.1 model (from Meta) has 126 layers.

Here’s a ridiculously oversimplified conceptual block diagram of a modern LLM.

There’s an initial stage which takes text and turns it into a sequence of vectors. Then, those sequence of vectors get passed through the layers in the middle. Finally, you get your answer out at the end. (Note: I’m deliberately omitting discussion about what actually happens in the stages depicted by the oval and the diamond above, because I want to focus here on the layers in the middle for this post. I’m not going to talk at all about concepts like tokens, embedding, attention blocks, and so on. If you’re interested in these sorts of details, I highly recommend the video But what is a GPT? Visual intro to Transformers by Grant Sanderson).

We can imagine the LLM as a system that recognizes patterns at different levels of abstraction. The first and last layers deal directly with representations of words, so they have to operate at the word level of abstraction, let’s think of that as the lowest layer. As we go deeper into the network initially, we can imagine each layer as dealing with patterns at a higher level of abstraction, we could call them concepts. Since the last layer deals with words again, layers towards the end would be at a lower layer of abstraction.

But, really, this talk of encoding patterns at increasing and decreasing levels of abstraction is all pure speculation on my part, there’s no empirical basis to this. In reality, we have no idea what sorts of patterns are encoded in the middle layers. Do they correspond to what we humans think of as concepts? We simply have no idea how to interpret the meaning of the vectors that are generated by the intermediate layers. Are the middle layers “higher level” than the outer layers in the sense that we understand that term? Who knows? We just know that we get good results.

The things we call models have different kinds of applications. We tend to think first of scientific models, which are models that give scientists insight into how the world works. Scientific models are a type of model, but not the only one. There are also engineering models, whose purpose is to accomplish some sort of task. A good example of an engineering model is a weather prediction model that tells us what the weather will be like this week. Another good example of an engineering model is SPICE, which electrical engineers use to simulate electronic circuits.

Perceptrons started out as a scientific model of the brain, but their real success has been as an engineering model. Modern LLMs contain within them feedforward neural networks, which are the intellectual descendants of Rosenblatt’s perceptrons. Some people even refer to these as multilayer perceptrons. But LLMs are not an engineering model that was designed to achieve a specific task, the way that weather models or circuits models do. Instead, these are models that were designed to predict the next word in a sentence, and it just so happens that if you build and train your model the right way, you can use it to perform cognitive tasks that it was not explicitly designed to do! Or, as Sean Goedecke put it in a recent blog post (emphasis mine)

Transformers work because (as it turns out) the structure of human language contains a functional model of the world. If you train a system to predict the next word in a sentence, you therefore get a system that “understands” how the world works at a surprisingly high level. All kinds of exciting capabilities fall out of that – long-term planning, human-like conversation, tool use, programming, and so on.

This is a deeply weird and surprising outcome about building a text prediction system. We’ve built text prediction systems before. Claude Shannon was writing about probability-based models of natural language back in the 1940s in his famous paper that gave birth to the field of information theory. But it’s not obvious that once these models got big enough, we’d get results like we’re getting today, where you could ask the model questions and get answers. At least, it’s not obvious to me.

In 2020, the linguistics researchers Emily Bender and Alexander Koller published a paper titled Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data. This is sometimes known as the octopus paper, because it contains a thought experiment about a hyper-intelligent octopus eavesdropping on a conversation between two English speakers by tapping into an undersea telecommunications cable, and how the octopus could never learn the meaning of English phrases through mere exposure. This seems to contradict Goedecke’s observation. They also note how research has demonstrated that humans are not capable of learning a new language through mere exposure to it (e.g., through TV or radio). But I think the primary thing this illustrates is how fundamentally different LLMs are from human brains, and how little we can learn about LLMs by making comparisons to humans. The architecture of an LLM is radically different from the architecture of a human brain, and the learning processes are also radically different. I don’t think a human could learn the structure of a new language by being exposed to a massive corpus and then trying to predict the next word. Our intuitions, which work well when dealing with humans, simply break down when we try to apply them to LLMs.

The late philosopher of mind Daniel Dennett proposed the concept of the intentional stance, as a perspective we take for predicting the behavior of things that we consider to be rational agents. To illustrate it, let’s contrast it with two other stances he mentions, the physical stance and the design stance. Consider the following three different scenarios, where you’re asked to make a prediction.

Scenario 1: Imagine that a child has rolled a ball up a long ramp which is at a 30 degree incline. I tell you that the ball is currently rolling up the ramp at 10 metres / second and ask you to predict what its speed will be one minute from now.

Scenario 2: Imagine a car is driving up a hill at a 10 degree incline. I tell you that the car is currently moving at a speed of 60 km/h, and that the driver has cruise control enabled, also set at 60 km/h. I ask you to predict the speed of the car one minute from now.

A car with cruise control enabled, driving uphill

Scenario 3: Imagine another car on a flat road that going at 50 km/h, and is about to enter an intersection, and the traffic light has just turned yellow. Another bit of information I give you: the driver is heading to an important job interview and is running late. Again, I ask you to predict the speed of the car one minute from now.

In the first scenario (ball rolling up a ramp), we can predict the ball’s future speed by treating it as a physics problem. This is what Dennett calls the physical stance.

In the second scenario (car with cruise control enabled), we view the car as an artifact that was designed to maintain its speed when cruise control is enabled. We can easily predict that its future speed will be 60 km/h. This is what Dennett calls the design stance. Here, we are using our knowledge that the car has been designed to behave in certain ways in order to predict how it will behave.

In the third scenario (driver running late who encounters a yellow light), we think about the intentions of the driver: they don’t want to be late for their interview, so we predict that they will accelerate through the intersection. We predict that the driver will accelerate through the intersection, and so we predict their future speed will be somewhere around 60 km/h. This is what Dennett calls the intentional stance. Here, we are using our knowledge of the desires and beliefs of the driver to predict what actions they will take.

Now, because LLMs have been designed to replicate human language, our instinct is to apply to the intentional stance to predict their behavior. It’s a kind of pareidolia, we’re seeing intentionality in a system that mimics human language output. Dennett was horrified by this.

But the design stance doesn’t really help us either, with LLMs. Yes, the design stance enables us to predict that an LLM-based chatbot will generate plausible-sounding answers to our questions, because that is what it was designed to do. But, beyond that, we can’t really reason about its behavior.

Generally, operational surprises are useful in teaching us how our system works by letting us observe circumstances in which it is pushed beyond its limits. For example, we might learn about a hidden limit somewhere in the system that we didn’t know about before. This is one of the advantages of doing incident reviews, and it’s also one of the reasons that psychologists study optical illusions. As Herb Simon put it in The Sciences of the Artificial, Only when [a bridge] has been overloaded do we learn the physical properties of the materials from which it is built.

However, when an LLM fails from our point of view by producing a plausible but incorrect answer to a question, this failure mode doesn’t give us any additional insight into how the LLM actually works. Because, in a real sense, that LLM is still successfully performing the task that it was designed to do: generate plausible-sounding answers. We aren’t capable of designing LLMs that only produce correct answers, we can only do plausible ones. And so we learn nothing about what we consider LLM failures, because the LLMs aren’t actually failing. They are doing exactly what they are designed to do.

Dijkstra never took a biology course

Simplicity is prerequisite for reliability. — Edsger W. Dijkstra

Think about a system whose reliability had significantly improved over some period of time. The first example that comes to my mind is commercial aviation, but I’d encourage you to think of a software system you’re familiar with, either as a user (e.g., Google, AWS) or as a maintainer of a system that’s gotten more reliable over time.

Think of a system where the reliability trend looks like this

Now, for the system you have thought about where its reliability increased over time, think about what the complexity trend looks like over time for that system. I’d wager you’d see a similar sort of trend.

My claim about what the complexity trend looks like over time

Now, in general, increases in complexity don’t lead to increases in reliability. In some cases, engineers make a deliberate decision to trade off reliability for new capabilities. The telephone system today is much less reliable than it was when I was younger. As someone who grew up in the 80s and 90s, the phone system was so reliable that it was shocking to pick up the phone and not hear a dial tone. We were more likely to experience a power failure than a telephony outage, and the phones still worked when the power was out! I don’t think we even knew the term “dropped call”. Connectivity issues with cell phones are much more common than they ever were with landlines. But this was a deliberate tradeoff: we gave up some reliability in order to have ubiquitous access to a phone.

Other times, the increase in complexity isn’t the product of an explicit tradeoff but rather an entropy-like effect of a system getting more difficult to deal with over time as it accretes changes. This scenario, the one that most people have in mind when they think about increasing complexity in their system, is synonymous with the idea of tech debt. With tech debt the increase in complexity makes the system less reliable, because the risk of making a breaking change in the system has increased. I started this blog post with a quote from Dijkstra about simplicity. Here’s another one, along the same lines, from C.A.R. Hoare’s Turing Award Lecture in 1980:

There are two ways of constructing a software design: One way is to make it so simple that there are obviously no deficiencies, and the other way is to make it so complicated that there are no obvious deficiencies. The first method is far more difficult.

What Dijkstra and Hoare are saying is: the easier a software system is to reason about, the more likely it is to be correct. And this is true: when you’re writing a program, the simpler the program is, the more likely that you are to get it right. However, as we scale up from individual programs to systems, this principle breaks down. Let’s see how that happens.

Djikstra claims simplicity is a prerequisite for reliability. According to Dijkstra, if we encounter a system that’s reliable, it must be a simple system, because simplicity is required to achieve reliability.

reliability ⇒ simplicity

The claim I’m making in this post is the exact opposite: systems that improve in reliability do so by adding features that improves reliability, but come at the cost of increased complexity.

reliability ⇒ complexity

Look at classic works on improving the reliability of real-world systems like Michael Nygard’s Release It!, Joe Armstrong’s Making reliable distributed systems in the presence of software errors, and Jim Gray’s Why Do Computers Stop and What Can Be Done About It? and think about the work that we do to make our software systems more reliable, functionality like retries, timeouts, sharding, failovers, rate limiting, back pressure, load shedding, autoscaling, circuit breakers, transactions, and auxiliary systems we have to support our reliability work like an observability stack. All of this stuff adds complexity.

Imagine if I took a working codebase and proposed deleting all of the lines of code that are involved in error handling. I’m very confident that this deletion of code would make the codebase simpler. There’s a reason that programming books tend to avoid error handling cases in their examples, they do increase complexity! But if you were maintaining a reliable software system, I don’t think you’d be happy with me if I submitted a pull request that deleted all of the error handling code.

Let’s look at the natural world, where biology provides us with endless examples of reliable systems. Evolution has designed survival machines that just keep on going; they can heal themselves in simply marvelous ways. We humans haven’t yet figured out how to design systems which can recover from the variety of problems that a living organism can. Simple, though, they are not. They are astonishingly, mind-boggling-y complex. Organisms are the paradigmatic example of complex adaptive systems. However complex you think biology is, it’s actually even more complex than that. Mother nature doesn’t care that humans struggle to understand her design work.

Now, I’m not arguing that this reliability-that-adds-complexity is a good thing. In fact, I’m the first person who will point out that this complexity in service of reliability creates novel risks by enabling new failure modes. What I’m arguing instead is that achieving reliability by pursuing simplicity is a mirage. Yes, we should pay down tech debt and simplify our systems by reducing accidental complexity: there are gains in reliability to be had through this simplifying work. But I’m also arguing that successful systems are always going to get more complex over time, and some of that complexity is due to work that improves reliability. Successful reliable systems are going to inevitably get more complex. Our job isn’t to reduce that complexity, it’s to get better at dealing with it.

The same incident never happens twice, but the patterns recur over and over

“No man ever steps in the same river twice. For it’s not the same river and he’s not the same man” – attributed to Heraclitus

After an incident happens, many people within the organization are worried about the same incident happening again. In one sense, the same incident can never really happen again, because the organization has changed since the incident has happened. Incident responders will almost certainly be more effective at dealing with a failure mode they’ve encountered recently than one they’re hitting for the first time.

In fairness, if the database falls over again, saying, “well, actually, it’s not the same incident as last time because we now have experience with the database falling over so we were able to recover more quickly” isn’t very reassuring to the organization. People are worried that there’s an imminent risk that remains unaddressed, and saying “it’s not the same incident as last time” doesn’t alleviate the concern that the risk has not been dealt with.

But I think that people tend to look at the wrong level of abstraction when they talk about addressing risks that were revealed by the last incident. They suffer from what I’ll call no-more-snow-goon-ism:

Calvin is focused on ensuring the last incident doesn’t happen again

Saturation is an example of a higher-level pattern that I never hear people talk about when focusing on eliminating incident recurrence. I will assert that saturation is an extremely common pattern in incidents: I’ve brought it up when writing about public incident writeups at Canva, Slack, OpenAI, Cloudflare, Uber, and Rogers. The reason you won’t hear people discuss saturation is because they are generally too focused on the specific saturation details of the last incident. But because there are so many resources you can run out of, there are many different possible saturation failure modes. You can exhaust CPU, memory, disk, threadpools, bandwidth, you can hit rate limits, you can even breach limits that you didn’t know existed and that aren’t exposed as metrics. It’s amazing how much different stuff there is that you can run out of.

My personal favorite pattern is unexpected behavior of a subsystem whose primary purpose was to improve reliability, and it’s one of the reasons I’m so bear-ish about the emphasis on corrective actions in incident reviews, but there are many other patterns you can identify. If you hit an expired certificate, you may think of “expired certificate” as the problem, but time-based behavior change is a more general pattern for that failure mode. And, of course, there’s the ever-present production pressure.

If you focus too narrowly on preventing the specific details of the last incident, you’ll fail to identify the more general patterns that will enable your future incidents. Under this narrow lens, all of your incidents will look like either recurrences of previous incidents (“the database fell over again!”) or will look like a completely novel and unrelated failure mode (“we hit an invisible rate limit with a vendor service!”). Without seeing the higher level patterns, you won’t understand how those very different looking incidents are actually more similar than you think.

Labeling a root cause is predicting the future, poorly

Why do we retrospect on our incidents? Why spend the time doing those write-ups and holding review meetings? We don’t do this work as some sort of intellectual exercise for amusement. Rather, we believe that if we spend the time to understand how the incident happened, we can use that insight to improve the system in general, and availability in particular. We improve availability by preventing incidents as well as reducing the impact of incidents that we are unable to prevent. This post-incident work should help us do both.

The typical approach to post-incident work is to do a root cause analysis (RCA). The idea of an RCA is to go beyond the surface-level symptoms to identify and address the underlying problems revealed by the incident. After all, it’s only by getting at the root at the problem that we will be able to permanently address it. When doing an RCA, when we attach the label root cause to something, we’re making a specific claim. That claim is: we should focus our attention on the issues that we’ve labeled “root cause”, because spending our time addressing these root causes will yield the largest improvements to future availability. Sure, it may be that there were a number of different factors involved in the incident, but we should focus on the root cause (or, sometimes, a small number of root causes), because those are the ones that really matter. Sure, the fact that Joe happened to be on PTO that day, and he’s normally the one that spots these sorts of these problems early, that’s interesting, but it isn’t the real root cause.

Remember that an RCA, like all post-incident work, is supposed to be about improving future outcomes. As a consequence, a claim about root cause is really a prediction about future incidents. It says that of all of the contributing factors to an incident, we are able to predict which factor is most likely to lead to an incident in the future. That’s quite a claim to make!

Here’s the thing, though. As our history of incidents teaches us over and over again, we aren’t able to predict how future incidents will happen. Sure, we can always tell a compelling story of why an incident happened, through the benefit of hindsight. But that somehow never translates into predictive power: we’re never able to tell a story about the next incident the way we can about the last one. After all, if we were as good at prediction as we are at hindsight, we wouldn’t have had that incident in the first place!

A good incident retrospective can reveal a surprisingly large number of different factors that contributed to the incident, providing signals for many different kinds of risks. So here’s my claim: there’s no way to know which of those factors is going to bite you next. You simply don’t possess a priori knowledge about which factors you should pay more attention to at the time of the incident retrospective, no matter what the vibes tell you. Zeroing in on a small number of factors will blind you to the role that the other factors might play in future incidents. Today’s “X wasn’t the root cause of incident A” could easily be tomorrow’s “X was the root cause of incident B”. Since you can’t predict which factors will play the most significant roles in future incidents, it’s best to cast as wide a net as possible. The more you identify, the more context you’ll have about the possible risks. Heck, maybe something that only played a minor role in this incident will be the trigger in the next one! There’s no way to know.

Even if you’re convinced that you can identify the real root cause of the last incident, it doesn’t actually matter. The last incident already happened, there’s no way to prevent it. What’s important is not the last incident, but the next one: we’re looking at the past only as a guide to help us improve in the future. And while I think incidents are inherently unpredictable, here’s a prediction I’m comfortable making: your next incident is going to be a surprise, just like your last one was, and the one before that. Don’t fool yourself into thinking otherwise.

On work processes and outcomes

Here’s a stylized model of work processes and outcomes. I’m going to call it “Model I”.

If you do work the right way, that is, follow the proper processes, then good things will happen. And, when we don’t, bad things happen. I work in the software world, so by “bad outcome” a mean an incident, and by “doing the right thing”, the work processes typically refer to software validation activities, such as reviewing pull requests, writing unit tests, manually testing in a staging environment. But it also includes work like adding checks in the code for unexpected inputs, ensuring you have an alert defined to catch problems, having someone else watching over your shoulder when you’re making a risky operational change, not deploying your production changes on a Friday, and so on. Do this stuff, and bad things won’t happen. Don’t do this stuff, and bad things will.

If you push someone who believes in this model, you can get them to concede that sometimes nothing bad happens even though someone didn’t do everything can quite right, the amended model looks like this:

Inevitably, an incident happens. At that point, we focus the post-incident efforts on identifying what went wrong with the work. What was the thing that was done wrong? Sometimes, this is individuals who weren’t following the process (deployed on a Friday afternoon!). Other times, the outcome of the incident investigation is a change in our work processes, because the incident has revealed a gap between “doing the right thing” and “our standard work processes”, so we adjust our work processes to close the gap. For example, maybe we now add an additional level of review and approval for certain types of changes.

Here’s an alternative stylized model of work processes and outcomes. I’m going to call it “Model II”.

Like our first model, this second model contains two categories of work processes. But the categories here are different. They are:

What people are officially supposed to
What people actually do

The first categorization is an idealized view of how the organization thinks that people should do their work. But people don’t actually do their work their way. The second category captures what the real work actually is.

This second model of work and outcomes has been embraced by a number of safety researchers. I deliberately called my models as Model I and Model II as a reference to Safety-I and Safety-II. Safety-II is a concept developed by the resilience engineering researcher Dr. Erik Hollnagel. The human factor experts Dr. Todd Conklin and Bob Edwards describe this alternate model using a black-line/blue-line diagram. Dr. Steven Shorrock refers to the first category as work-as-prescribed, and the second category as work-as-done. In our stylized model, all outcomes come from this second category of work, because it’s the only one that captures the actual work that leads to any of the outcomes. (In Shorrock’s more accurate model, the two categories of work overlap, but bear with me here).

This model makes some very different assumptions about the nature of how incidents happen! In particular, it leads to very different sorts of questions.

The first model is more popular because it’s more intuitive: when bad things happen, it’s because we did things the wrong way, and that’s when we look back in hindsight to identify what those wrong ways were. The second model requires us to think more about the more common case when incidents don’t happen. After all, we measure our availability in 9s, which means the overwhelming majority of the time, bad outcomes aren’t happening. Hence, Hollnagel encourages us to spend more time examining the common case of things going right.

Because our second model assumes that what people actually do usually leads to good outcomes, it will lead to different sorts of questions after an incident, such as:

What does normal work look like?
How is it that this normal work typically leads to successful outcomes?
What was different in this case (the incident) compared to typical cases?

Note that this second model doesn’t imply that we should always just keep doing things the same way we always do. But it does imply that we should be humble in enforcing changes to the way work is done, because the way that work is done today actually leads to good outcomes most of the time. If you don’t understand how things normally work well, you won’t see how your intervention might make things worse. Just because your last incident was triggered by a Friday deploy doesn’t mean that banning Friday deploys will lead to better outcomes. You might actually end up making things worse.

When a bad analysis is worse than none at all

One of the most famous physics experiments in modern history is the double-split experiment, originally performed by the English physicist Thomas Young back in 1801. You probably learned about this experiment in a high school physics class. There was a long debate in physics about whether light was a particle or a wave, and Young’s experiment provided support for the wave theory. (Today, we recognize that light has a dual nature, with both particle-like and wave-like behaviors).

To run the experiment, you need an opaque board that has two slits cut out of it, as well as a screen. You shine a light at the board and look to see what the pattern of light looks like on the screen behind it.

Here’s a diagram from Wikipedia, which shows the experiment being run with electrons rather than light , but is otherwise the same idea.

Original: NekoJaNekoJa Vector: Johannes Kalliauer, CC BY-SA 4.0 https://creativecommons.org/licenses/by-sa/4.0, via Wikimedia Commons

If light was a particle, then you would expect each light particle to pass through either one slit, or the other. The intensities that you’d observe on the screen would look like the sum of the intensities if you ran the experiment by covering up one slit, and then ran it again by covering up the other slit. It should basically look like the sum of two Gaussian distributions with different means.

However, that isn’t what you actually see on the screen. Instead, you get this pattern where there are some areas of the screen with no intensity at all: where the light never strikes the screen. On the other hand, if you run the experiment by covering up either slit, you will get light at these null locations. This shows that there’s an interference effect, the fact that there are two slits leads the light to behave differently from being the sum of the effects of each slit.

Note that we see the same behavior with electrons (hence the diagram above). Both electrons and light (photons) exhibit this sort of wavelike behavior. This behavior is observed even even if you shine only one electron (or photon) at a time through the slits.

Now, imagine a physicist in the 1970s hires a technician to run this experiment with electrons. The physicist asks the tech to fire one electron at a time from an electron gun at the double-slit board, and record the intensities of the electrons striking a phosphor screen, like on a cathode ray tube (kids, ask your parents about TVs in the old days). Imagine that the physicist doesn’t tell the technicians anything about the theory being tested, the technician is just asked to record the measurements.

Let’s imagine this thought process from the technician:

It’s a lot of work to record the measurements from the phosphor screen, and all of this intensity data is pretty noisy anyways. Instead, why don’t I just identify the one location on the screen that was the brightest, use that location to estimate which slit the electron was most likely to have passed through, and then just record that slit? This will drastically reduce the effort required for each experiment. Plus, the resulting data will be a lot simpler to aggregate than the distribution of messy intensities from each experiment.

The data that the technician records then ends up looking like this:

Experiment	Slit
1	left
2	left
3	right
4	left
5	right
6	left

Now, the experimental data above will give you no insight into the wave nature of electrons, no matter many experiments are run. This sort of experiment is clearly not better than nothing, it’s worse than nothing, because it obscures the nature of the phenomenon that you’re trying to study!

Now, here’s my claim: when people say “the root cause analysis process may not be perfect, but it’s better than nothing”, this is what I worry about. They are making implicit assumptions about a model of incident failure (there’s a root cause), and the information that they are capturing about the incidents is determined by this model.

A root cause analysis approach will never provide insight into how incidents arise through complex interactions, because it intentionally discards the data that could provide that insight. It’s like the technician who does not record all of the intensity measurements, and instead just uses those measurements to pick a slit, and only records the slit.

The alternative is to collect a much richer set of data from each incident. That more detailed data collection is going to be a lot more effort, and a lot messier. It’s going to involve recording details about people’s subjective observations and fuzzy memories, and it will depend on what types of questions are asked of the responders. It will also depend on what sorts of data you even have available to capture. And there will be many subjective decisions about what data to record and what to leave out.

But if your goal is to actually get insights from your incidents about how they’re happening, then that effortful, messy data collection will reveal insights that you won’t ever get from a root cause analysis. Whereas, if you continue to rely on root cause analysis, you are going to be misled about how your system actually fails and how it really works. This is what I mean by good models protect us from bad models, and how root cause analysis can actually be worse than nothing.

Don’t be like the technician, discarding the messy data because it’s cleaner to record which slit the electron went through. Because then you’ll miss that the electron is somehow going through both.

You can’t prevent your last outage, no matter how hard you try

I don’t know anything about your organization, dear reader, but I’m willing to bet that the amount of time and attention your organization spends on post-incident work is a function of the severity of the incidents. That is, your org will spend more post-incident effort on a SEV0 incident compared to a SEV1, which in turn will get more effort than a SEV2 incident, and so on.

This is a rational strategy if post-incident effort could retroactively prevent an incident. SEV0s are worse than SEV1s by definition, so if we could prevent that SEV0 from happening by spending effort after it happens, then we should do so. But no amount of post-incident effort will change the past and stop the incident from happening. So that can’t be what’s actually happening.

Instead, this behavior means that people are making an assumption about the relationship between past and future incidents, one that nobody ever says out loud but everyone implicitly subscribes to. The assumption is that post-incident effort for higher severity incidents is likely to have a greater impact on future availability than post-incident effort for lower severity incidents. In other words, an engineering-hour of SEV1 post-incident work is more likely to improve future availability than an engineering-hour of SEV2 post-incident work. Improvement in future availability refers to either prevention of future incidents, or reduction of the impact of future incidents (e.g., reduction in blast radius, quicker detection, quicker mitigation).

Now, the idea that post-incident work from higher-severity incidents has greater impact than post-incident work from lower-severity incidents is a reasonable theory, as far as theories go. But I don’t believe the empirical data actually supports this theory. I’ve written before about examples of high severity incidents that were not preceded by related high-severity incidents. My claim is that if you look at your highest severity incidents, you’ll find that they generally don’t resemble your previous high-severity incidents. Now, I’m in the no root cause camp, so I believe that each incident is due to a collection of factors that happened to interact.

But don’t take my word for it, take a look at your own incident data. When you have your next high-severity incident, take a look at N high-severity incidents that preceded it (say, N=3), and think about how useful the post-incident incident work of those previous incidents actually was in helping you to deal with the one that just happened. That earlier post-incident work clearly didn’t prevent this incident. Which of the action items, if any, helped with mitigating this incident? Why or why not? Did those other incidents teach you anything about this incident, or was this one just completely different from those? On the other hand, were there sources of information other than high-severity incidents that could have provided insights?

I think we’re all aligned that the goal of post-incident work should be in reducing the risks associated with future incidents. But the idea that the highest ROI for risk reduction work is in the highest severity incidents is not a fact, it’s a hypothesis that simply isn’t supported by data. There are many potential channels for gathering signals of risk, and some of them come from lower severity incidents, and some of them come from data sources other than incidents. Our attention budget is finite, so we need to be judicious about where we spend our time investigating signals. We need to figure out which threads to pull on that will reveal the most insights. But the proposition that the severity of an incident is a proxy for the signal quality of future risk is like the proposition that heavier objects fall faster than lighter one. It’s intuitively obvious; it just so happens to also be false.

Good models protect us from bad models

One of the criticisms leveled at resilience engineering is that the insights that the field generates aren’t actionable: “OK, let’s say you’re right, that complex systems are never perfectly understood, they’re always changing, they generate unexpected interactions, and that these properties explain why incidents happen. That doesn’t tell me what I should do about it!”

And it’s true; I can talk generally about the value of improving expertise so that we’re better able to handle incidents. But I can’t take the model of incidents that I’ve built based on my knowledge of resilience engineering and turn that into a specific software project that you can build and deploy that will eliminate a class of incidents.

But even if these insights aren’t actionable, that they don’t tell us about a single thing we can do or build to help improve reliability, my claim here is that these insights still have value. That’s because we as humans need models to make sense of the world, and if we don’t use good-but-not-actionable models, we can end up with actionable-but-not-good models. Or, as the statistics professor Andrew Gelman put it in his post The social sciences are useless. So why do we study them? Here’s a good reason back in 2021:

The baseball analyst Bill James once said that the alternative to good statistics is not no statistics, it’s bad statistics. Similarly, the alternative to good social science is not no social science, it’s bad social science.

The reason we do social science is because bad social science is being promulgated 24/7, all year long, all over the world. And bad social science can do damage.

Because we humans need models to make sense of the world, incidents models are inevitable. A good-but-not-actionable incident model will feel unsatisfying to people who are looking to leverage these models to take clear action. And it’s all too easy to build not-good-but-actionable models of how incidents happen. Just pick something that you can measure and that you theoretically have control over. The most common example of such a model is the one I’ll call “incidents happen because people don’t follow the processes that they are supposed to.” It’s easy to call out process violations in incident writeups, and it’s easy to define interventions to more strictly enforce processes, such as through automation.

In other words, good-but-not-actionable models protect us from the actionable-but-not-good models. They serve as a kind of vaccine, inoculating us from the neat, plausible, and wrong solutions that H.L. Mencken warned us about.

Tradeoff costs in communication

If you work in software, and I say the word server to you, which do you think I mean?

Software that responds to requests (e.g., http server)
A physical piece of hardware (e.g., a box that sits in a rack in a data center)
A virtual machine (e.g., an EC2 instance)

The answer, of course, is it depends on the context. The term server could mean any of those things. The term is ambiguous; it’s overloaded to mean different things in different contexts.

Another example of an overloaded term is service. From the end user’s perspective, the service is the system they interact with:

From the end user’s perspective, there is a single service

But if we zoom in on that box labeled service, it might be implemented by a collection of software components, where each component is also referred to as a service. This is sometimes referred to as a service-oriented architecture or a microservice architecture

A single “service” may be implemented by multiple “services”. What does “service” mean here?

Amusingly, when I worked at Netflix, people referred to microservices as “services”, but people also referred to all of Netflix as “the service”. For example, instead of saying, “What are you currently watching on Netflix?”, a person would say, “What are you currently watching on the service?”

Yet another example is the term “client”. This could refer to the device that the end-user is using (e.g., web browser, mobile app):

Or it could refer to the caller service in a microservice architecture:

It could also refer to the code in the caller service that is responsible for making the request, typically packaged as a client library.

The fact that the meaning of these terms is ambiguous and context-dependent makes it harder to understand what someone is talking about when the term is used. While the person speaking knows exactly what sense of server, service or client they mean, the person hearing it does not.

The ambiguous meaning of these terms creates all sorts of problems, especially when communicating across different teams, where the meaning of a term used by the client team of a service may be different from the meaning of the term used by the owner of a service. I’m willing to bet that you, dear reader, have experienced this problem at some point in the past when reading an internal tech doc or trying to parse the meaning of a particular slack message.

As someone who is interested in incidents, I’m acutely aware of the problematic nature of ambiguous language during incidents, where communication and coordination play an enormous role in effective incident handling. But it’s not just an issue for incident handling. For example, Eric Evans advocates the use of ubiquitous language in software design. He pushes for consistent use of terms across different stakeholders to reduce misunderstandings.

In principle, we could all just decide to use more precise terminology. This would make it easier for listeners to understand the intent of speakers, and would reduce the likelihood of problems that stem from misunderstandings. At some level, this is the role that technical jargon plays. But client, server and service are technical jargon, and they’re still ambiguous. So, why don’t we just use even more precise language?

The problem is that expressing ourselves in unambiguous isn’t free: it costs the speaker additional effort to be more precise. As a trivial example, microservice is more precise than service, but it takes twice as long as to say, and it takes an additional five letters to write. Those extra syllables and letters are a cost to the speaker. And, all other things being equal, people prefer expending less effort than more effort. This is why we don’t like being on the receiving end of ambiguous language, because we have to put more effort into resolving the ambiguity through context clues.

The cost of precision to the speaker is clear in the world of computer programming. Traditional programming languages require an extremely high degree of precision on behalf of the coder. This sets a very high bar for being able to write programs. On the other hand, modern generative AI tools are able to take natural language inputs as specifications which are orders of magnitude less precise, and turn them into programs. These tools are able to process ambiguous inputs in ways that regular programming languages simply cannot. The cost in effort is much lower for the vibe programmer. (I will leave evaluation of the outcomes of vibe programming as an exercise for the reader).

Ultimately, the degree in precision in communication is a tradeoff: an increase in precision means less effort for the listener and less risk of misunderstanding, at a cost of more effort for the speaker. Because of this tradeoff, we shouldn’t expect the equilibrium point to be at maximal precision. Instead, it’s somewhere in the middle. Ideally, it would be where we minimize the total effort. Now, I’m not a cognitive scientist, but this is a theory that has been advanced by cognitive scientists. For example, see the paper The communicative function of ambiguity in language by Steven T. Piantadosi, Harry Tily, and Edward Gibson. I touched on the topic of ambiguity more generally in a previous post the high cost of low ambiguity.

We often ask “why are people doing X instead of the obviously superior Y“. This is an example of how we are likely missing the additional costs of choosing to do Y over X. Just because we don’t notice those costs doesn’t mean they aren’t there. It means we aren’t looking closely enough.

Model error

One of the topics I wrote about in my last post was about using formal methods to build a model of how our software behaves. In this post, I want to explore how the software we write itself contains models: models of how the world behaves.

The most obvious area is in our database schemas. These schemas enable us to digitally encode information about some aspect of the world that our software cares about. Heck, we even used to refer to this encoding of information into schemas as data models. Relational modeling is extremely flexible: in principle, we can represent just about any aspect of the world into it, if we put enough effort in. The challenge is that the world is messy, and this messiness significantly increases the effort required to build more complete models. Because we often don’t even recognize the degree of messiness the real world contains, we build over-simplified models that are too neat. This is how we end up with issues like the ones captured in Patrick McKenzie’s essay Falsehoods Programmers Believe About Names. There’s a whole book-length meditation on the messiness of real data and how it poses challenges for database modeling: Data and Reality by William Kent, which is highly recommended by Hillel Wayne, in his post Why You Should Read “Data and Reality”.

The problem of missing the messiness of the real world is not at all unique to software engineers. For example, see Christopher Alexander’s A City Is Not a Tree for a critique of urban planners’ overly simplified view of human interactions in urban environments. For a more expansive lament, check out James C. Scott’s excellent book Seeing Like a State. But, since I’m a software engineer and not an urban planner or a civil servant, I’m going to stick to the software side of things here.

Models in the back, models in the front

In particularly, my own software background is in the back-end/platform/infrastructure space. In this space, the software we write frequently implement control systems. It’s no coincidence that both cybernetics and kubernetes derived their names from the same ancient Greek word: κυβερνήτης. Every control system must contain within it a model of the system that it controls. Or, as Roger C. Conant and W. Ross Ashby put it, every good regulator of a system must be a model of that system.

Things get even more complex on the front-end side of the software world. This world must bridge the software world with the human world. In the context of Richard Cook’s framing in Above the Line, Below the Line, the front-end is the line that bridges the two world. As a consequence, the front-end’s responsibility is to expose a model of the software’s internal state to the user. This means that the front-end also has an implicit model of the users themselves. In the paper Cognitive Systems Engineering: New wine in new bottles, Erik Hollnagel and David Woods referred to this model as the image of the operator.

The dangers of the wrongness of models

There’s an oft-repeated quote by the statistician George E.P. Box: “All models are wrong but some are useful”. It’s a true statement, but one that focuses only on the upside of wrong models, the fact that some of them are useful. There’s also a downside to the fact that all models are wrong: the wrongness of these models can have drastic consequences.

And, while It’s a true statement, but what it fails to hint at how bad the consequences can be when a model is wrong. One of my favorite examples involves the 2008 financial crisis, as detailed by the journalist Felix Salmon’s 2009 Wired Magazine article Recipe for Disaster: The Formula that Killed Wall Street. The article described how Wall Street quants used a mathematical model known as the Gaussian copula function to estimate risk. It was a useful model that ultimately led to disaster.

Here’s A ripped-from-the-headlines example of image of the operator model error, how the U.S. national security advisor Mike Waltz accidentally saved the phone number of Jeffrey Goldberg, editor of the Atlantic magazine, to the contact information of White House spokesman Brian Hughes. The source is the recent Guardian story How the Atlantic’s Jeffrey Goldberg got added to the White House Signal group chat:

According to three people briefed on the internal investigation, Goldberg had emailed the campaign about a story that criticized Trump for his attitude towards wounded service members. To push back against the story, the campaign enlisted the help of Waltz, their national security surrogate.

Goldberg’s email was forwarded to then Trump spokesperson Brian Hughes, who then copied and pasted the content of the email – including the signature block with Goldberg’s phone number – into a text message that he sent to Waltz, so that he could be briefed on the forthcoming story.

Waltz did not ultimately call Goldberg, the people said, but in an extraordinary twist, inadvertently ended up saving Goldberg’s number in his iPhone – under the contact card for Hughes, now the spokesperson for the national security council.

…

According to the White House, the number was erroneously saved during a “contact suggestion update” by Waltz’s iPhone, which one person described as the function where an iPhone algorithm adds a previously unknown number to an existing contact that it detects may be related.

The software assumed that, when you receive a text from someone with a phone number and email address, that the phone number and email address belong to the sender. This is a model of the user that turned out to be very, very wrong.

Nobody expects model error

Software incidents involve model errors in one way or another, whether it’s an incorrect model of the system being controlled, an incorrect image of the operator, or a combination of the two.

And, yet, despite us all intoning “all models are wrong, some models are useful”, we don’t internalize that our systems our built on top of imperfect models. This is one of the ironies of AI: we are now all aware of the risks associated with model error with LLMs. We’ve even come up with a separate term for it: hallucinations. But traditional software is just as vulnerable to model error as LLMs are, because our software is always built on top of models that are guaranteed to be incomplete.

You’re probably familiar with the term black swan, popularized by the acerbic public intellectual Nassim Nicholas Taleb. While his first book, Fooled by Randomness, was a success, it was the publication of The Black Swan that made Taleb a household name, and introduced the public to the concept of black swans. While the term black swan was novel, the idea it referred to was not. Back in the 1980s, the researcher Zvi Lanir used a different term: fundamental surprise. Here’s an excerpt of a Richard Cook lecture on the 1999 Tokaimura nuclear accident where he talks about this sort of surprise (skip to the 45 minute mark).

And this Tokaimura accident was an impossible accident.

There’s an old joke about the creator of the first English American dictionary, Noah Webster … coming home to his house and finding his wife in bed with another man. And she says to him, as he walks in the door, she says, “You’ve surprised me”. And he says, “Madam, you have astonished me”.

The difference was that she of course knew what was going on, and so she could be surprised by him. But he was astonished. He had never considered this as a possibility.

And the Tokaimura was an astonishment or what some, what Zev Lanir and others have called a fundamental surprise which means a surprise that is fundamental in the sense that until you actually see it, you cannot believe that it is possible. It’s one of those “I can’t believe this has happened”. Not, “Oh, I always knew this was a possibility and I’ve never seen it before” like your first case of malignant hyperthermia, if you’re a an anesthesiologist or something like that. It’s where you see something that you just didn’t believe was possible. Some people would call it the Black Swan.

Black swans, astonishment, fundamental surprise, these are all synonyms for model error.

And these sorts of surprises are going to keep happening to us, because our models are always wrong. The question is: in the wake of the next incident, will we learn to recognize that fundamental surprises will keep happening to us in the future? Or will we simply patch up the exposed problems in our existing models and move on?