Epistemic Slipperiness
(epistemic status: thinking out loud, kinda confused)
Sometimes I interact with a person, and something about their thinking feels… slippery.
Probably, sometimes my mind is slippery. But, I can’t quite pin down when.
This is probably actually several different phenomena, but they blur together in my current mind. I’m confused about it. This post is me trying to deconfuse myself. Let’s explore a few examples and see what shakes out.
Subtly Bad Jokes and Slipping Sideways
I first got the concept of ‘epistemic slipping’ from this Eliezer facebook post in 2017, which argued that there was something off about people who made jokes about how Trump winning the presidency meant we must be living in a simulation.
When I’m asking whether some person known to be in the Advanced Epistemology 101 club is *strongly* trustworthy, one of the primary questions I ask is: Does this person slip sideways in reality, even just a little, in order to resolve their internal tensions?
You don’t know in advance what crisis is going to hit. But when a crisis hits, it often creates some mental tension that can be diminished by slipping sideways in reality to a world that is false but not *blatantly* false. No, person Y couldn’t really be lying to me!
And since I see this sideways-slipping as this huge remaining foundational sin even for many people in the AE101 club… well, you can see why I was worried about hearing some people making Simulation Hypothesis jokes about Trump. It expresses a wish for this world not to be real, an impulse to slip sideways out of the tension, even if only by saying that this world is a simulation and so it doesn’t count. The conceptual link to the Simulation Hypothesis seems to show a forming fault-line—like the person’s brain now has a new way to escape from reality in case of crisis. I worry both that the general pressure exists and hasn’t been defeated, and that the Simulation Hypothesis in particular is a fault-line that could crack if the person gets put under enough pressure.
(ADDED: It’s not that the fault-line says a Trump victory is okay because simulations are ontologically unreal. The feeling is rather that simulations don’t count toward the Laws that Must Be Obeyed, the socially-valent generalizations whose violation feels uncomfortable. It was mandatory for Hillary Clinton to win in the real world, the world that *counts*; but she doesn’t have to win in a simulated world. That’s why the joke is funny.)
The concept of ‘slipping sideways’ seemed really important to me, and seemed to come up in multiple other places. More recently, I’ve felt a sense of slipperiness as I watch some people argue about various x-risk related topics. And I sometimes see people make plans that don’t make sense, in ways that are hard to pin down.
Lately I’ve been wanting to get more clarity on this for a few reasons:
First, I probably slip sideways sometimes, and I want to get better at noticing that, and not doing so.
Second, I want to notice when other people are slippery, so that I can trust them ‘an appropriate amount’, which may mean filtering them out of some hiring processes.
Third, ideally when I notice someone being epistemically slippery I’d like to communicate with them about it.
Okay, what are some other examples
I can’t actually think of examples that have exactly the same quality as the Trump/Simulation thing. But when I query for examples of people who have felt ‘slippery’ in some way to me in the past, I get:
The ideologue who doesn’t notice themselves shifting goalposts, imposing doublestandards, or generally notice that they’re “fighting for a side” rather than truthseeking.
The young idealist who wants to “make a difference”, but ends up trying to justify a strategy to themselves that seems to be a confused mess of “make use of the college degree they already started”, “feel vaguely good about themselves”, “be prestigious”, “make their parents proud”. Their plan is sorta reasonable, but when you try to ask them about what they’re actually trying to achieve, their reasoning feels subtly off. If you ask “why are you doing X instead of Y?” they somehow… slide off the question.
The startup founder who’s really excited about their plan/mission and keeps talking about it. And their plan… doesn’t really make sense, and somehow manages to make it seem like each new fact supports their existing mission even if it feels like it should be falsifying it.
The angry guy conflating things, who is upset at Alice for doing what he perceived as violation of an important norm. But he doesn’t separate out his observations about Alice from his judgments and assumptions about Alice, and then presents that mishmash as if it were an objective truth.
The manipulative negotiator who is deliberately vague about what they’re offering and what benefits you’ll receive, and when you try to pin them down they say things that superficially sound like they’re saying something concrete, but when you go back and check the record you realize they were still kinda weaseling out of committing to anything.
The manipulative guru who keeps responding your concerns with vaguely reasonable true-ish sounding things, that nonetheless leave you feeling painted into a corner.
The guy who’s really stuck in their ontology, where you to make an argument, but it doesn’t fit into their frame and they keep rounding the things you’re trying to say to either the nearest concept in their own frame. Or, they conclude you’re talking nonsense and don’t bother listening.
The guy who’s stuck in their ontology (while trying to solve their own problems). Same as previous, but even when they’re really motivated to solve their own problem, without arguing with anyone (or maybe talking to a close friend they’re earnest trying to listen to), they still keep missing concepts that don’t fit into their preconceptions.
The strangely disinterested guy, who you tell something pretty important that you think should be really relevant to their worldview (but maybe implies they’re missing something important), and they… just don’t seem that interested.
Why call it “Slipperiness?”
Okay, so I listed a bunch of examples. Are they actually the same phenonemon? The examples contain a bunch of rationality-errors, but maybe those are just a bunch of disparate errors and “epistemic slipperiness” is just “not being very good at rationality in a bunch of unrelated domains.”
Eliezer’s example was specifically about resolving internal mental pressures. Is that what all of them are about? Is that sufficient or useful?
A Noticing-Handle
I think I ended up writing this post because I’ve observed a bunch of situations that ‘felt slippery’ to me, and ‘feeling slippery’ was a noticing-handle I could explore, and then build a habit off of. (i.e. ‘notice feeling of slipperiness’ ->> ‘do something on purpose to reduce the problems of the slipperiness’)
Upon reflection, I think ‘slipperiness’ is mostly a feeling I get around other people.
When I’m slipping, the feeling is more specific. When I’ve been the ideologue, or idealist with the confused plan, or the guy stuck-in-his-ontology, those each feel differently from the inside. It feels like defensiveness, or fear, or righteously doing the right thing.
But when I’m talking to someone else, and I keep trying to point at some reasonable concept, and they keep missing it, and then persist in missing even after I’ve tried to point it out multiple ways, that gives me a slippery sensation.
Taxonomy of Slipperiness
Some distinctions that come up when I look over each of the above examples, and reflect on where slipperiness has come up in my life.
Invisible rationality errors that you completely fail to notice.
Persistent, reinforced rationality errors that maybe you do notice, or someone points them out, but then even when it’s pointed out you still have trouble seeing it.
Communication failures. These usually involve at least one person making some kind of “internal rationality-error”, but I think there’s then some additional failures-that-feel-slippery where one person doesn’t quite know how to communicate, and another person doesn’t know how to listen.
I think these are mostly different phenomena.
For me, the “trump/simulation” thing was invisible, but not persistent. When I read the FB post, I was like “oh, I totally might have made the trump/simulation joke, and not noticed that it was pointing at an important sideways slide into non-reality, and yes, upon reflection it totally was an important sideways slide to avoid living in an uncomfortable reality.” But, as soon as it was pointed out, I was like “oh, yeah that makes sense.”
By contrast – I’ve sometimes been “the angry guy conflating observations with assumptions.” And then someone points that out, and then I’m either like “nu-uh” or like “I dunno, maybe, but you’re being unreasonable” and not really hearing the thing. I’m in the middle of an argument, and much of my cognition is geared towards “winning”. I’m incentivized not to notice the rationality-error that I’m making. And it’s not until I’m outside the current conflict that I’m able to look at it clearly and thoughtfully. The failure is persistent.
Finally, there are moments I perceive ‘slipperiness’, where the main ingredient seems to be a joint-failure-to-communicate. This particularly comes up with frame/ontology disagreements, where one person is trying to “communicate in their native language”, and the other person just has a totally different native language.
Depth in persistent slipperiness.
One type of slipping is to kinda notice the type of error you’re making, and then convince yourself that you now “get it”, when in fact you only got a pale shadow of the error you’re making.
Imagine getting into a fight with your spouse. They’re upset that you got them the wrong kind of birthday cake. You might think “oh I get it, they really care about cake. Sorry for messing that up.” But, actually, what they’re upset about is that they mentioned their preferred cake flavor recently, and you didn’t pay attention and remember. Wait, no, actually the problem is that you’ve been not paying that much attention to their preferences for months/years, and the cake situation just happened to be the straw that broke the camel’s back.
And meanwhile you have a self-image as a good spouse, so each step along the way you’re got an incentive to think to yourself “ah, I get it, I’m paying attention and doing a good job listening”, when in fact the problem was subtler and required more active listening.
Here, the cognitive error is “skipping to the step where you’re pretty sure you understand the problem and are ready to execute on solution, before listening and fleshing out that understanding, and then doing that multiple times.” And it’s noteworthy that you can successfully gain a new taste for part of the problem (“spouse has strong opinions about cake”), which indeed will successfully predict some future problems (indeed, now whenever the topic of cake comes up they’re kinda tense because it reminds them of the last time you didn’t listen to them and now it’s kinda become a Whole Thing)
Appendix
I have a more thoughts on slipperiness, but they feel less “blogpost” shaped and more like “random thoughts” shaped. A teaser of some things I’m still thinking about:
How to talk to someone who seems to persistently be missing concepts, in a slippery way? In default-culture this is kinda insulting to state directly.
I mentioned “frame confusion” as a way you could get a “slippery communication failure.” But I also think “frame confusion” might be a source of slipperiness inside a single person. The Young Idealist With a Confused Plan might be equivocating between different types of “successful plans”, and ways of thinking about successful plans, and not noticing when they’ve made the switch.
Coordination and “fake plans.” Part of my motivation here is a certain kind of slipperiness that shows up in people’s x-risk plans, that I’m not sure what to do about.
I feel like I recognize this feeling of perceived slipperiness, but the way I experience it does not match all the examples. From the list, the Angry person and both variants of Ontology person feel like bouncing off. This isn’t a case of them sort-of-trying to engage the problem and having a low-effort method of eluding uncomfortable truths; they are putting additional effort into rejecting the idea outright or changing it into something else more agreeable.
Following Eliezer’s example, compare two different reactions to the wrong team winning an election:
Haha, we’re living in a simulation.
They did not win. The election was stolen, and the correct team’s candidate is the true winner. Everything the government does is illegal and I don’t have to listen to it.
The first is slippery; the second is not.
It feels to me like a distinguishing characteristic of the slipperiness is the attempt to grasp the thing-which-slips. This helps to explain the frequency with which the slippery feeling and confusion appear together: confusion is a hagfish.
More degrees of freedom in a representation give it more ways to map between things (transfer learning), but more degrees of freedom are the opposite of what makes for predictive outputs.
If you can retroactively fit a utility function to any sequence of actions, what predictive power do we gain by including utility functions into our models of AGI?
Reading this I get the sense that slipperiness is special cases of favoring your priors too strongly in one way or another. I point this out because it seems like an instance of the general class of issues I might call “confusion” where you fail to notice what the world is telling you and you believe something because it fits how you see the world rather than believing what the world is showing you.
I thought about confusion as I was writing this post. Confusion shares the property with “slipperiness” of being hard-to-notice. I currently think it’s a different thing, although I agree there’s some similarities.
I think of confusion as a specific phenonema (whereas “Slipperiness” is more of a description of a phenomena). In my language, confusion is when you don’t have an explanation for what’s going on. (Failing to notice you’re confused is an additional thing that can happen to you. Maybe failing to notice that you’re confused is a “slippery” state)
My model is of this kind of cognitive failures is “motivated reasoning due to imprecise binding”.
(Though I like “slipperiness”, it’s nice and pithy).
My model is that the “outer edges” of words/concepts/heuristics are relatively large. The binding between A and B is not precise. The joint has slack in it.
And within that “box”, the outer edges of seeming reasonableness, there is some kind of internal heuristic or emotional-agent that is trying to “save face” or to make itself feel better or preserve its current belief system, using what slack or degrees of freedom it has to accomplish that, while typically remaining below the radar of conscious awareness.
I’m not sure if this explanation makes sense. Perhaps I am, in part, talking about a different kind of cognitive error.
I often notice “the person who really likes her iconoclast boyfriend, and defends him in public interactions about him by claiming her partner believes something more normal than he in fact believes.” (pronouns for less ambiguous grammar and because I think the gender usually skews that way)
If I physically think of slippiness then its about the object being near where I would have predicted to be but not quite. Maybe it could be due to ignoring air resistance or static friction being different from gracing friction. Those factors would seem that they probably are independent from each other. But they share the property in that I am not accounting for them.
So if one knows what correct reasoning looks like one can detect when a thing is not quite it. It might be tempting to focus on the “correct way” of doing reasoning if one has a great wish to bring more of it. But this can interfere with modelling thoughts of those that are doing it “incorrectly”. So if there is a unexplained force it might be worth it to ask what is the force. This might lead to being frustrated with the limitations with others (and I guess self). And because it doesn’t result in vulgar errors it might be tempting to keep the reference to correct reasoning. Because its “only slippery” then engaging in a “pure” way should still get you somewhere around the correct area. But at the same time at some step the approximation might break to an outrigth flaw.
If you are waling on a bridge and it creaks you have the option to get curious why it creaks. And it might be important to notice that you don’t know why it creaks. You also have the option to get habituated and accept that it creaks. If you are in a hurry to go from A to B ignorance might prodive you with speed. If you don’t model it if it collapses it is goning to do so unacceptedly. And it might not be possible to deduce from the creak how the bridge needs to be repaired (you might actually need to look at support beams etc).
I’m not sure if this is relevant, since “slipperiness” is only an analogy, but: I don’t normally think of slipperiness as a modeling failure, but as a control failure.
“Slipperiness” evokes the sort of problem where you are applying sufficient force to accomplish your goal, but your goal is not being achieved, because your force is being subtly redirected into an unintended direction (hence, the soap moves sideways instead of up).
This doesn’t necessarily involve surprise; maybe I thought I only had a 10% chance of successfully picking up the soap, so the soap slipping sideways was always the expected outcome. You might reason “if you knew it wasn’t going to work, why did you do it?” but it could be that I just don’t have a better strategy available.
However, it does feel significant that many examples from the post involve someone being unaware that their attempt at rationality has failed. That implies both a failure of control and a subsequent failure of perception. (Then again, that second problem seems to crop up all over the place in epistemology, so maybe it’s not specifically related to the “slipperiness” thing.)
I would suppose that if one had better model of traction control then soaps would not slip from hands. Jugglers are doing grabbing very accurately. Trying to grab 3 objects at once as a non-juggler will result in one of the objects slipping away largely because the object tracking is overwhelming.
I think my model for soap is approximately as detailed as my models for rocks or sticks, but the soap remains harder to pick up. That seems like a fact about the soap, rather than about my model of the soap.
If one is slipping on a ice road its too much speed for the driving conditions. On the other hand its possible to drift even on hot dry asphalt. A stun driver might burn rubber but have centimeter control to make the car go exactly where he wants. Intoxicated people slip more for standard walking tests (an DUI has different people have different control levels on the same environment so that seems to be a property of the drivers rather than the roads). You can also break from run to a stop by sliding on your two feet (for example on a gravel road). When you do this you are not often called to be “slipping”.
Also freudian slips are called slips. Usually the formed words are perfectly understandable ie fully pronounced. A lot of slip words are not challenging on the basis of their letters. And for example japanese people slipping more on english R and L is not really a property of the letters themselfs as other nationalities “grip” on them more strongly.
Extending the analog made me realise that “optimal gripping” that John Vaervake is so much a fanboy of, fits the pattern. And in that vein of memes, having arguments whether a feature is subjective or objective can be confusing if the phenomenon in question is transjective. Whether a being can effectively relate to a feature of the world depends upon the shape of both. “Slip” in this picture can be thought of as ineffectualness or wastefulness in the relatedness, an inferior stance for which a better relationship exists and is known.
I realize this is not the most important point but man I love that your name is Slider and you’re having this convo.
I no longer believe that I understand what point you are trying to make, but a few remarks that each seem relevant to at least one of those examples:
Slipperiness is a continuum, not a binary trait.
Slipperiness is just one factor that affects your probability of success. For instance, a skilled ice-skater has less chance of falling down than a newbie, but a given patch of ice is equally “slippery” for both. (As I use the word.)
I could buy that slipperiness is a function of two surfaces interacting, rather than something that a single surface can unambiguously be said to possess. (Though knowing just one of the inputs seems to provide a lot of information about the output; e.g. wet soap is unusually slippery when combined with lots of other surfaces, not just a few.)
All of the above seem to me to be fully compatible with my original model of slipperiness as a control problem rather than a modeling problem. Climbing a steep hill is harder than climbing a gentle slope for reasons that have nothing to do with epistemology. Climbing a slippery hill is also harder for reasons that have nothing to do with epistemology.
You think it is more important to emphasise the control problem and I think that modelling is better / there is no need to priotise control. Its going to be coupled because control will utilise feedback so its hard to tease appart.
In principle I could look at the sole of my shoe with a microscope and do the same for the ground. However often I decide not to gather this information (which is a quite sensible decision) and choose to be ignorant about the microstructures. Rather I use a quick sketch of guess of the statistical properties. The cost of the ignorance is that reality does take the details into account, so in that aspect I lose (partially) track of reality. Unexpected behaviour is when that model leakage becomes apparent to me.
With the iceskating it provides a very clear example of a clear surface between metal and ice burshing against each other. This is where slipperyness should shine but if the skater retains control we don’t call this slipping. There is also the phenomenon where the ice unexpectedly lacking in slipperiness can make a skater trip (slips due to stickyness have their own unique name). I am really not exited about the concepts that would asssign “inherent difficulty” to parts of terrain so easy/hard not a fan of (stuff like friction coefficients are real though). With ice for example a skater can travel unhindered in smooth ice and has trouble covering bumpy ice while a pedestrian maintains easy access in bumpy ice and slips on smooth ice. This is a counterexample on newbie vs pro iceskater.
Hmm, “slipperiness” sounds like a concept that’s more useful for observing my thoughts, rather than a way to think about my interactions with others.
I can tell when my mind is trying to gloss over something (e.g. when I don’t feel like specifying my entire chain of thought, then when I do try writing out a proper explanation, I start seeing loopholes and finding counter examples. Or when I dismiss something too quickly or for no reason.).
However, if I sense that someone’s not fully engaging with my points, how do I know if that’s because I’ve misunderstood them, or because they have knowledge they’re unable to articulate, or because there’s motivated reasoning at play etc.?
If I can’t tell, then I’d think it makes sense to treat it as a general case of communication failure instead of a specific case of “slipperiness”, i.e. I would try to understand what they are saying, find and ask about the apparent contradiction stated from their perspective, and observe their response.
If I do manage to identify it as one of the “slipperiness” cases, then different examples would require different approaches (e.g. disengage for manipulators, give space to calm down for emotional outbursts), so it wouldn’t make sense to treat them as a general “slipperiness” case?
I think that consumption of fiction can be a symptom of epistemic slipping. Fiction can feel more wholesome or comfortable or pleasing than the real world. Joy in the Merely Real is an epistemic habit.