Almost true
08 June 2026
Listen to my narration of this piece (10 min)
Recently, I’ve been thinking a lot about things that are almost true, specifically about other people and myself.
Modern deep learning techniques have uncovered an interesting phenomenon, in that on the one hand they are extremely good at image recognition, i.e. the problem of: is this a picture of a cat or is it a picture of a dog? And yet we find that for almost any technique, even state-of-the-art techniques, the models that are learned can be fooled rather easily. You can add a very small perturbation to the image, i.e. you change the value of each pixel by a negligible amount, and it can go from being very confident that it’s a cat to very confident that it’s a dog. The image has changed in no discernible way to the human eye. We’re talking about very, very small changes to each pixel value, and this is interesting because what it’s saying is that to the model, this picture of a cat is very close in some direction to it being a picture of a dog. But that’s not true of the way that we as humans understand images and understand our visual sensory input. Visual illusions are, though they exist, fairly rare and they’re kind of exceptional when they occur. I’m thinking about things like M.C. Escher’s continuous stairs, or the colour of the dress: whether it’s blue and black or white and gold. So what does this tell us? This tells us that we’re actually very good at recognising objects, and this detection moreover is very robust in a way that we haven’t been able to reproduce with machine learning techniques. This, despite us having produced some extraordinarily accurate techniques. And I should say that the perturbations that are used to fool these models are in some way pathological. It’s a very particular perturbation that we’re making. Though it’s very small, it is very particular. It’s not necessarily something that you would see in the wider world, or at least it’s exceedingly unlikely, which is the thing that these models care about (it hasn’t shown up in the training dataset). There is much to be discussed on that topic, so look up adversarial perturbations in machine learning if interested. Related but not the same is the concept of the uncanny valley, which loosely is the phenomena that things which appear very close but not the same as humans (a wax figure) are more disturbing and evoke more negative reaction than things which are like-but-not-too-like a human (a classic sci-fi robot).
Why have I been thinking about this? Well, it occurs to me that this idea that you’ve got something in front of you and there is some sort of accurate depiction of it, and yet we might be fooled or it might immediately look like something else, is relevant more widely. We’re very good at it for object detection, as I say, but I don’t think we’re very good at it for personalities. There might be a good reason for this. I think that evolutionarily it’s plausible that is was super important to correctly distinguish between edible plants and poisonous plants, as well as perhaps needing to be able to tell the difference between an animal in the distance that is something we may predate and an animal that will predate us. These things are important and have been important for our survival, and it could be that it’s not been so important historically that we are good at detecting the personalities of people. I expect for a long time we didn’t get as much choice as we do now about who is in our lives, and so this discernment didn’t really make a difference. Add to this the fact that for personality traits you get very little data, and the problem of identifying personality based on first impression becomes difficult. I think there is a lot to be said for intuition in this regard, but that is a topic for another day.
It’s an interesting thing to reflect on, from an introspective point of view. What is it about me that I think is true but that someone else viewing me could think is false? What is it of my attributes that could easily and reasonably be interpreted or be taken as something else? I’ve thought about this quite a lot because it’s always very interesting to me when I come across new people. Besides learning about them, you get to observe them meeting you for the first time. And if you forget your ego, and the idea of wanting to look good, then you can really get a lot out of this. This person is meeting you for the first time and, though you may have an idea for yourself that’s baked in, this person will likely have a very different idea. Their reaction will tell you a lot. Consider the following questions. What might someone think when they’re meeting you for the first time? What do you look like at first glance? How does this change as someone grows to know you? Sometimes these things are strikingly different from your own personal identity. I’ll give an example from my perspective, because I have recently gained a small piece of understanding which has cleared up some things for me.
I categorise myself as a very curious person. I ask lots of questions when I’m interacting with someone for the first time. I think that’s not just because I’m interested in lots of different things - but I would say this is true - but also that I usually acquire interest in whatever the other person is interested in. I just become very curious about what that person is saying. But this paired with my mathematical training and, I guess, analytical framework for thinking manifests itself as pointed questions and really digging down into the details of whatever we are talking about. This can come across as judgemental, particularly if the conversation is about a specific belief the other person has. The reason it can come across like this, is that it would look pretty much exactly the same if I were judgemental. I would be asking lots of pointed questions, I would be focused on their answers, I would be slightly pedantic. So from my perspective, I (believe I) am doing it out of curiosity - not because I want to know the truth even - and to understand the perspective of the other person. But this requires quite a lot of questioning and sometimes quite foundational questioning that can be quite uncomfortable for the person. I’ve come to learn that this thing that I think about myself, which is that I’m very non-judgemental, and driven by a pluralistic, relativistic view of the world meaning I find it very easy to hold the position that contradictory opinions can and should coexist, can in fact be interpreted as the opposite. My interactions with a person can easily be construed - I don’t say misconstrued because … well … because of my relativistic view of the world - as judgemental. It’s an interesting exercise to reflect on what it is about yourself that you hold quite dearly but that people tend not to think about you at least to first impression. You can think about this with respect to yourself you can also think about it with respect to your friends. I find that it’s a nice conversation to have with people that know you quite well, as well as people that don’t know you very well - to explore what it is about yourselves individually that as a group you think are true but that a new person could reasonably think otherwise. I have found that this exercise helps with interacting with new people because you understand what that person might be seeing. Sometimes we get frustrated because the person that we’re talking to for the first time isn’t seeing the person that we believe we are, but having this conversation allows you to understand that view of yourself and perhaps mitigate it in certain ways. I now know that I can come across as judgmental in the first regard, and it doesn’t necessarily change how I am in the conversation - I think this can lead to self-consciousness and inauthenticity - but it does mean that I’m not taken by surprise.
As soon as you make this realisation, one implication is that you can be far more generous to those you are meeting. Just as those who meet you might easily get the wrong idea, you might get the wrong idea of others.
So, what’s almost true about you?