theinexactsciences.github.io

1.5

Suspended,

It feels somewhat strange to respond to a letter that was not addressed to me, especially that is now perhaps more of a historical artefact. Nor was your letter purloined, but simply made public, published even, as part of a correspondence. In that sense, this is perhaps more of a commentary, or simply a piece of fan mail.

The main reason I felt compelled to respond was your conceptualization of reading as inference, which is very close to how I have been thinking about related issues.

In particular, the example that kept coming to mind, while I read your letter, is that of “stotting”. This is a behavior that has been observed in animals such as deer, in which they needlessly jump while running from a predator. Since this sort of jumping does not increase the speed at which they can flee from the predator (and may in fact slow them down or make them more visible), this behaviour has been interpreted by some of those who study it as a kind of signaling. In effect, those animals who stot are flaunting their ability, suggesting to the predator that it should not bother trying to chase them down, because they clearly have enough speed and stamina to outrun the danger. The predator, in turn, will read this signaling appropriately, redirecting its efforts towards an animal that is not indicating that it is capable of escaping, which is more likely to be one that it will be able to catch (perhaps because the prey is old, or injured, for example).

The key to this, and the only reason it works, as others have noted, is that this signal is hard to fake. It wouldn’t work for animals to try to convey their ability to outrun a predator using some means that was easily available to all; if all individuals are equally capable of sending that signal, no matter what their ability to actually outrun a predator, then anti-inductivity would kick in, and all of them would do so (at which point, the message would lose all meaning, effectively conveying no information, so in fact none of them would). It is only because the signal really does represent something that is highly indicative of ability to run fast that it works as a signal. Those who are not fast enough to stot will get eaten.

The same sort of thing of course explains many strange behaviours among animals, as far as I understand, including things like bird plumage and mating behaviours (we can also think of obvious analogies among humans). That does not mean, of course, that all such signals retain their intrinsic meaning forever. Especially among humans, evolutionary psychologists have suggested that all kinds of preferences can be traced to things which were once signals representing various characteristics likely to be beneficial for offspring, such as the ability to have children, or overall health or freedom from disease. These are now basically evolutionary holdovers because a) many of these are no longer so hard to fake, and b) they are no longer strong indicators of the relevant outcomes. Nevertheless, there is of course in which they can continue to have power, in so far things that are desirable themselves become symbols of status, in a self-reinforcing feedback loop.

Whether or not the above explanation is correct (let’s say for the sake of argument that it is), we might conventionally think of the observation and reaction by the predator as a kind of “reading”, though it instinctively seems less appropriate to call the stotting itself a kind of “writing”. The first explanation we might instinctively provide for this difference is that stotting is not a conscious, calculated choice by the prey (“if I leap now, then that wolf will infer that I am strong, and realize how foolish it would be to chase me”), but rather behaviour that is essentially automatic, like some part of the fight or (in this case) flight response. That seems reasonable, except that presumably the “reading” by the predator is just as similarly unconscious and automatic (if that is indeed a fair characterization of the prey’s behaviour).

Even though this seems slightly inconsistent, I think it can actually be reasonably understood within your framework. If we think of reading as a kind of inference, then the predator is accurately inferring from the prey’s stotting that it is a fast animal (assuming that the stotting is in fact a trustworthy signal), and adjusting its behaviour as a result. It obviously has not carried out a precise statistical inference, but rather has applied various heuristics to do a kind of approximate inference, incorporating the new information into its model of each deer’s ability and it’s likelihood of catching each of them. The deer is writing in a sense, signaling as it leaps, but “writing” still feels inappropriate. Perhaps the key is that it does not seem to be an individualized signal crafted to the particular situation, but just part of an automatic behaviour that would be triggered by any predator, and is similar across all prey, such that we would call it something more like “instinct”.

All that is to say, we can think of both reading and writing as being more or less calculated or automatic. All of these notions clearly apply to humanity as well, with the added complexity that we humans have much richer models of the world, much more cognitive capacity, and often much more flexibility to craft our signals or carry out our readings over an extended period of time.

That’s not to say, however, that we don’t also have more automatic forms of these. A classic example (again from evolutionary theory), might be something like blushing. Blushing seems to be one of those evolutionary puzzles for which we can come up with numerous possible explanations. Part of the thinking on this will be guided by salient features of blushing, specifically the fact that it is automatic and uncontrollable, and also (perhaps to a lesser extent), that it is hard to fake. One could perhaps learn to blush on command, or to avoid blushing in any circumstance, but I expect it would take considerable practice in either case. I won’t speculate here on the “true purpose” of blushing, but rather note that it a kind of writing that we carry out without intending it (to the extent that stotting is), and one that we similarly read in each other without needing to think about it; we see someone blush and immediately infer that they are embarrassed (or some more complex interpretation, depending on the context).

At the opposite extreme, I can finally turn to your discussion of more literal reading and writing, as with literature. Here there seems to be much more room for a more deliberate, calculated, drawn out form of intention and interpretation, though I think it is also more complicated than that. Especially on the writing side, it is no doubt the case that many times we very deliberately carry out actions that we believe will cause a reader to make particular inferences, and in some cases the calculated nature of this might sometimes lead this to be labeled manipulation.

To some extent this is what writers do. In editing a piece of work, a writer might think about how a particular sentence will be received, whether it is confusing, how it could be improved, based largely on how they think it will be read. At the same time, it seems implausible that writers begin purely with such calculated modeling. Rather, they begin. Either through practice, raw talent, or experience gained through feedback, they have some intuitive sense of what will work, and follow those intuitions, or even begin in the most non-deliberate writing they can, hoping that things will emerge, or that they will be able to shape the material into something later. One might also point to “automatic writing” as a tool used by writers to try to get past blocks or generate ideas, though ironically one seems to need to use fairly deliberate techniques in order to produce work that is adequately “automatic”.

On the reading side, there is again a continuum. Even professional critics presumably have many automatic reactions while they are reading, though these of course will be shaped by their experience, and will likely be quite different from that of a more naive reader. (The critic is far less likely to be affected by simple manipulative techniques like music inserted to trigger emotional cues, and much more likely to have a negative reaction to overused tropes). Nevertheless, the interpretive part of a literal kind of reading (or listening or viewing) will likely be a more extended activity, with interpretations developed through reflection, discussion, debate, etc.

The scope for such reading is limitless of course. One very narrow type of reading is to assume that there is a single specific “meaning” intended by the writer, and that the goal is to figure out what that is. I’m not familiar with Knapp & Michaels, but based on your description, this seems to be the only type of meaning they are interested in, which seems to me to be an extremely impoverished view.

Certainly there are some settings in which that view is relevant. For example, in creating a crossword puzzle, the creator has an intended answer for each clue, such that the correct answers will cohere. The goal of the puzzle solver is largely to infer what the writer intended for each clue, although in rare circumstances, there could be multiple solutions which would fit equally well, in which case one might fail to infer the intended meaning, and yet still “solve” the puzzle. (Ironically, puzzle solving can be turned into puzzle creation, with enough creativity in interpreting the clues).

Another example would be something like detecting what seem to be typos, and trying to infer what the author intended to write, if they did in fact intend to write something different. In reading your letter, I noticed you say “in a class narrow-and-conquer method”, which doesn’t quite scan. I thought at first that you meant to write “crass”, but upon reflection, you probably intended “classic”. Perhaps it was both!

As for the “meaning” of most art, however, it seems overly simplistic to suggest that there is a single “intended” meaning by the creator, for the reasons discussed above. There might be ambiguities in the text for which the writer has a personal interpretation, but in most cases the larger “meaning” will not be a single intention by the author (except perhaps in their intention to create a particular emotional or affective response to a particular passage, etc.).

Moreover, there is simply much more than can be inferred. Indeed, much of the elaboration of criticism is in thinking about art as information which can be used to make inferences about the world, either about the world in which the work was created, or about aspects of the author which they themselves may not even be aware of, at least not consciously.

One subtlety I might suggest is worth drawing out a bit more is your suggestion that “inference is performed statistically”. This is clearly true in the sense that we are dealing with uncertainties. We are not making syllogistic deductions, but rather assembling evidence into models. The reason I find this interesting is that we do now have very good theories and even systems for performing “correct” statistical inference. In fact, I was particularly struck (although I didn’t notice it on my first read through) by your mention of a quest for a “so-called universal method of statistical inference which can be uniformly, automatically applied to any problem”. This should be the subject for another letter, but there is a sense in which this is something that has in fact now been theorized and created by the statistical community.

The conclusion from such statistical theorizing and building, however, seems to be that this is not something that humans could plausibly be carrying out (because anything exact would be intractable). Hence my reference above to heuristics, and perhaps the reason that we can still be so easily fooled by crude signals. This also connects to your characterization of the link between color and capacity for action (within the context of traffic signals) as statistical, though I won’t try to develop that here.

Regardless, reading will of course be dependent on our models of the world, and can be manipulated by misunderstanding. You mention GPT-3, which provides many such examples. In the absence of any prompting to that effect, many people would likely assume that various texts by GPT-3 were written by a person, and might proceed to interpret based on that assumption. Moreover, such interpretations might be “correct” within the confines of their assumptions. Upon learning that such text was in fact produced by an anthropomorphized computer program, many would similarly assume that the machine must in some sense be conscious or otherwise capable of various feats. Again, such an inference is not necessarily wrong within the confines of its assumption, though that is only because most people would be wrong about what they believe is possible or impossible to do with systems trained on large amounts of text.

In that sense, the ability to write coherently is something that, up until now, has been quite hard to fake. Because it must be learned, writing was a clear signal of a certain level of cognitive ability, and other abilities that could reliably be expected to come along with that. Now that GPT-3 exists, of course, people will need to recalibrate, though it seems likely that most people will remain perpetually behind in making truly accurate inferences, given the pace at which things are changing. Nevertheless, even if we do successfully adapt our conscious machinery in interpreting texts, it seems highly likely that we will continue to be “fooled” to some extent by convincing text. Just as we can’t help but blush, and can’t help but read blushing, we likely can’t help but read text as meaningful, in the sense of having had a consciousness behind it, even if we then follow that with a more rigorous or creative reading.

Possible Modernist