Developing frames

Note: I wrote this post in 2023, along with the rest of my meta-rationality sequence. I never got around to uploading it, but a recent comment from Vladimir Nesov inspired me come back to it and fill in the few remaining gaps. I still broadly stand by the ideas in this post (some of which were also covered in my posts on Why I'm Not a Bayesian and Towards a Formal Scientific Epistemology), though I don't necessarily still endorse all the specific claims. My recent thinking along these lines tries to be much more precise, and no longer uses the "frame" terminology very much. Previously in this sequence I described frames as clusters of mental traits. But what does that look like in practice? In this post I’ll explore in more detail how frames are structured, and how that affects the ways we learn and improve them. The structure of frames One reason for using the “frame” terminology is that frames can be seen as an attempt to address the frame problem in AI and philosophy: the problem of how we identify which knowledge is relevant to a particular situation. In order to orient ourselves to a given context, we need to identify what to pay attention to, how to interpret it, and what affordances we have in response. This often occurs not via explicit reasoning, but rather via a type of perceptual shift, as in the famous duck-rabbit illusion. Once we snap into seeing the image as a duck or a rabbit, we can identify the most salient features—there’s the eye, there’s the mouth, and so on—and interpret the rest in light of those. One way of making sense of such shifts is in terms of the predictive processing framework. According to this framework, our perceptions aren’t just determined from the bottom up by sensory inputs, but also from the top down by our expectations. More specifically, predictive processing postulates a hierarchy of generative models which both pass signals up to the models above them, and make predictions about the models below them. If strong enough, these predictions can even override the signals coming upwards. A top-down prediction of “rabbit” or “duck” can therefore force a reinterpretation of features identified at lower levels. Frames do the same thing at a more abstract level. (Indeed, it seems plausible that frames should be formalized as hierarchical generative models; I haven’t attempted this primarily because I’ve lacked the time to flesh out the details.) While the duck-rabbit image is an example of the same low-level sensory inputs being interpreted in different ways, a more abstract version of this involves the same high-level concepts being interpreted in different ways. Some examples: "Electricity" means very different things to a quantum physicist, an electrical engineer, a biologist, a meteorologist, and a layperson—while in some sense they're all thinking about the same concept, their frames focus on very different aspects of it, and there may be few or no specific claims about electricity which are relevant to all their frames. Consider recounting the history of the 20th century from the perspective of different ideologies: each one will tell an overarching story referring to many of the same events, but with the overall narratives being very different. People used to classify animals and plants based on their physical features; in the biological frame, though, these same features are used for categorization primarily via being evidence about their location on the evolutionary tree. Consider looking at the same house from the frame of a builder vs an architect vs a real estate agent vs an economist vs an artist vs a city planner. Each of them will have the concepts of windows and doors and many others in common, but they will focus on different features, and assign them different significance. Here’s an updated version of the diagram from the previous post, which includes concepts being reused across frames. However, while I’ve focused on concepts for the sake of simplicity, I previously characterized frames as being composed of a much wider range of mental phenomena, which are also reused across different frames. For example, habits and heuristics like seeking out more information or being cooperative might apply in different ways in different scenarios; A scientist might have a set of deep-rooted intuitions about how to set up, carry out and interpret experiments, which they’re unable to communicate explicitly. (And indeed, over the last few decades philosophers of science have shifted towards characterizing even science as a practice where the empirical components of theories need to be understood in the context of a range of background assumptions, which should be studied from an ethnological or anthropological lens.) The most general version of meta-rationality would treat each of these concepts as frames in their own right, for example as portrayed in the diagram below (with arrows omitted for simplicity). Learning and improving frames Aside from the ones we learn