Models

A model provides a grammar and a vocabulary. When it 'works' we think there is some additional similarity, some isomorphism, which underlies this. All there is is a judgement - and we cannot speculate on whether such judgements are, in general, reliable.

Talking about talking

We can imagine a language which doesn't have certain self-reflective capacities - e.g. which doesn't have an explicit truth predicate. This is because we would still be able to translate this language into the one we are presently speaking.

(We can't, of course, translate - or, therefore, imagine - a language which did have a truth predicate, but whose truth predicate worked in a significantly different way from 'ours'.)

Some language-like games might not be able to have a truth predicate. The introduction of a truth predicate (which would have to, again, be isomorphic with 'our' truth predicate) would render them unintelligible. Bargaining games are like this - when a salesman swears that everything in the brochure is true, we know he cannot be using 'is true' in the same way as, say, a logician or a natural scientist. On the other hand, the bargaining game would be unplayable if it was complemented by a logical or natural-science-style truth predicate, because this game depends upon a certain amount of conventional dishonesty. The extent of this dishonesty is explored by the participants in the game, but not explicitly explored - it is explored practically, by finding out what moves 'work' and what moves don't. Salesman do not make good philosophers, but this does not render the bargaining game intrinsically corrupt.

What would render it corrupt would be dishonest practice on the part of any players. I'm going to give a definition of this here which I'm not sure I can fully support, but which I think is roughly right:

Dishonest practice is practice which exploits the expectations of other players in a way which sacrifices the playability of the game to the instrumental aims of the practitioner.

Since the bargaining game cannot include a truth predicate, and since it is hard to construct a normatively self-reflective game without a truth predicate, it is possible that the bargaining game cannot be used to explictly adjudiate on honest practice. To test any practice against the definition, we would need to make a judgement about the objects and consequences of moves within the game - and particualrly about whethe a move rendered the game unplayable. This last, I think, would definitely require self-consious reflection on the possibility of truth-telling.

This raises a tricky moral issue: Can we, from outside, using a language with the capacity to articulate truth related issues, adjudicate on the language of the bargainers, which does not? We would be presuming to be able to translate this language, while knowing that we could not discuss the quality of this translation with the bargainers - we could not ask them whether the translation was correct.

We could, of course, learn to play the game. If we had an appropriate facility with it, we could then reflect on this facility using our native meta-language. This looks tricky: the character of our engagement would be different just because we had access to the meta-language. A physicist discussing mass with a weights and measures inspector does not participate in the weights-and-measures game in the same way as another inspector would.

This may be harmless. We certainly wouldn't want to say that the physicist and the weights and measures inspector didn't understand one another - we wouldn't want to say that they weren't able to play this game together. We would, however, find that their game became more difficult if it developed in a certain self-reflective way. There would be some truths about mass that they could only share by changing the game.

Natural Laws

We can't just say that it is a law of nature that nature follows laws.

What kind of law can this law be?

We can only talk about a predictable nature - anything else would be incomprehensible. Natural laws are the grammar of this talk.

Conviction, Certainty and Argument

The world is careless of our phenomenological convictions - it produces them without warranting them. We might imagine that we have some 'evolutionary guarantee' that the internal products of our sensory apparatus contribute to our survival, but this comforting mythology can't be grounded, itself, in our phenomenological convictions; and Darwinian evolution selects on the basis of publicly available behaviours and characteristics - not on the normative character of intentional states. It is indifferent between lucky misconception and 'true knowledge'.

We might argue that, over time, evolution would tend to weed out ignorance and misconception. There isn't any evidence for this beyond our own, present, existence. This existence, though, is the outcome of just one of many rolls of the evolutionary dice; and we know already that it is very probably temporary. Attributing intentional states to past 'evolutionary successes' in support of a phenomenological evolutionary epistemology would be circular and gratuitous.

It is absurd to say that our whole language game is reliable because it has 'survived' an evolutionary process which can only be given an intelligible account of if the game is reliable ...

Martian intentionality

And we couldn't unambiguously attribute intentional states to them - such as knowing how to tell the truth in Human.

Goodman, chess machines, and physical standards

It's as though we decided to mean by 'green', 'parrot coloured'. Now a previously grue parrot would remain green, but the meaning of green would change.

Chess playing computers

While we might learn a lot about chess from the behaviour of a computer, and might learn how to play from such a computer, we cannot define chess in terms of the computer's behaviour. The rules of chess determine whether the computer is playing correctly - they are not tested against the behaviour of the computer.

This is also true of people: we might believe that people act rationally, but we cannot define rationality in terms of the behaviour of any individual or group - however apparently exemplary.

This extends to their internal 'behaviour' - no neurological account of brain processes can capture the normative aspects of rationality, just as no actual 'or' gate could count as the effective standard for the logical operator.

Under certain circumstances, we might have to choose between deciding that the standard meter had grown or that our prior measurements all needed to be proportionally corrected. If we standardise the meaning of 'or' on the behaviour of an 'or' gate, the meaning of 'chess' on the behaviour of a computer, or the meaning of 'rationality' on either a neurological process or the behaviour of an individual or group, we do not know what adjustment this might lead us to have to make - we do not know what changes in meaning we might have to accomodate.

The standard metre can only change in one way, resulting in a uniform scalar adjustement to length. The standard 'or' gate could change the meaning of 'or' in a way which rendered the world unintelligible.

The standard chess computer might redefine chess in terms of any option available to it - whether or not it was functioning correctly.

Rules and Laws

So I think this is where I am:

Language requires attribution of intentional states, and is 'a priori' possible or we wouldn't be able to say so.

Certain irreducible intentional states - 'thoughts' (Frege?) - cannot be expressed unless it is possible to attribute generic properties. Predication requires universals.

To know how to talk is, among other things, to know how to attribute these generic properties.

Also - 'X knows how to talk' attributes some specific generic properties, including the ability to follow rules, to X.

To be able to attribute universal properties is part of being able to talk, it is part of being able to follow the rules of talk.

Asking whether the world 'really' contains universals (including whether it 'really' follows general laws); or asking whether certain people 'really' follow rules is like asking whether it is possible to talk about the world, or whether these people do, in fact, talk. It's a Moorean error - a question which can only have one answer if it is to make any sense at all.

To talk about a world is to talk about a world of this kind.

Rules and Intentionality

Various people (Crispin Wright, Ruth Millikan & others) have drawn attention to a potential role for intentionality in resolving Kripke's paradox, since intentional states seem to have the scope and reliability we would wish to attribute to competent rule-following.

I think it would be a mistake to take this (as Millikan and Wright do) as a cue for producing a better account of intentionality, if by 'account' we mean 'reductive account'. For reasons that I've probably laboured elsewhere in this blog (more open questions ...), I think any reductive account is almost certain to generate a version of Kripke's paradox.

An account of intentionality in terms of physical behaviour (e.g. of a machine) is obviously ruled out by this kind of objection, since questions about the reliability of the machine would always make sense - and these questions would be questions about whether the machine appropriately fulfilled its function; whether it appropriately 'modelled' the intentional state it was meant to underpin.

It is difficult to think of how a reductive account could be given that could avoid this machine model objection.

In concert with accounts of truth and meaning (to which intentionality is obviously related), a recursive account can, however, be given. Since we have to be able to attribute intentional states in order to be able to speak, and since we can speak, we can start with the intentional states we must be able to attribute in order to be able to have this conversation (we don't need a complete list - just an articulable selection).

Following from this, we can give an account of rule following in terms of these intentional states, or from others that depend upon them. To follow a rule is to act in accordance with an intentional attribution.

Intentionality and rules

If I am in a certain intentional state - say believing something particular to be true (that today is Friday), then I will in general behave and speak in a ways that is consistent with this. The ways that might be consistent cannot be spelled out in advance, and not only because they are complex - also because unexpected situations will arise.

Who will adjudicate? I and my interlocutors, playing the shared language game within which the statement of my intentional state is held to be true. The intelligibility, the playability, of these adjudications will be one part of what makes the whole game possible; and if the become idiosyncratically unintelligible then so does the original statement of my intentional state - we will find that we did not know, after all, what we meant by attributing this state to me.

This process of adjudication and exploration is what the history of mathematics reveals to us about addition. Someone who has been quadding so far, but thinks they were adding, is not a competent participant in this conversation.

Language Machines

What we actually say to one another is only philosophically interesting as a guide to what we can say to one another. And the limits of what we can say are not the limits of what we can think, from an internal, phenomenal perspective - although they do determine the limits of the thoughts we can attribute to one another. Further, the thoughts we can attribute to an interlocutor within the conversation we are having with him or her are more constrained than the thoughts we can attribute to someone with whom we are not having a conversation - or even cannot have a conversation. We can only incorrigibly attribute thoughts to an interlocutor (corrolary to Kripke).

I can say of someone 'he believes it is impossible to talk', but I cannot say to you 'you believe it is impossible to talk'. And 'I believe it is impossible to talk' directly undermines its own meaning, so must be false if it means anything. My attribution of this thought to someone is always corrigible, however.

We cannot attribute thoughts without attributing rule-following. (I think this is what Frege believed). And there are other reasons why rule following and intentionality go hand in hand. One is (again) to do with Kripke's paradox - we cannot define rule following in terms of behavioural descriptions which do not, themselves, refer to rules. This can be extended: we cannot define a rule in terms of the behaviour of any actual mechanical system. We cannot define 'or' in terms of the behaviour of some specific 'or' gate (a physical logical standard equivalent to the standard metre) without having some guarantee that the gate would never fail - and this guarantee obviously couldn't depend on the behaviour of some further physical system. What would count as failure is an irreducibly normative, and not simply a mechanical, issue.

To someone of a certain physicalist bent, this will sound wrong: after all, are we not, ourselves, physical sytems? How can we speak to one another if our speaking to one another requires the attribution of intentionality and rule following, while agreeing that rule following cannot be rendered physically?

This would be a reasonable question if it could be posed from outside our linguistic system (independently of all questioning and answering), but no quesiton can be posed from that standpoint. The most abstract, self-referential questions are still moves within the game. And if we cannot play the game without following rules then we must be able to follow rules - regardless of the clash with physicalist intuitions.

A more reductive, rather than reductio, response to the physicalist is to point out that physicalism can only be given substance by attributing rule based behaviour to 'nature', and this cannot be an outcome of grounded discovery. When we state a physical law we say that some things 'always' or 'everywhere' (given appropriate context) behave in some way (and, of course, the scope of any hypothetical law is itself, stated unconditionally: 'this only happens here' is not local even if 'this' is). Counterexamples, infamously, can be dealt with semantically as well as hypothetically. We discover and define as we go along - we find out a property, and then use it as a test (as in the boiling point of a liquid). We render the world intelligible by showing how it can be described - by discovering how we can talk about it, and, therefore, what thoughts about it we can attribute to interlocutors. (I might, as usual, add 'honest and competent' to 'interlocutor' except that these adjectives also directly determine the applicability of the descriptor - what would a dishonest and incompetent interlocutor look like?)

In other words, we can only construct the physicalist metaphor if we can talk to one another, and we can only do it by rendering the world in a way which appears, already, to contain semantic elements. This does not make it a bad metaphor, but it does make it an irreducible one - it can only ever be a heuristic, not an epistemological or metaphysical fundamental.

'We can talk about the world' is a fact about the world, but it cannot be given an 'account' of in terms of some other 'facts' that do not, already, depend upon it.

It is clear, in one sense, from all of this, that we can think of things that we cannot make. We cannot make a machine sufficiently reliable to be a standard for 'or', for instance; though we can know how 'or' works and, therefore, how such a machine should work if it could be constructed. Our best machines of this kind (the ones we can almost completely rely on) depend on aspects of our world that are most semantically secure. We would hardly know what physics was if we couldn't rely on the mechanisms from which our computers are built - and by this, I mean that we would regard someone who questioned these mechanisms as asking unintelligible questions. To have them break down would be like having an intelligent friend suddenly begin to talk nonsense. We could have no conversation with this friend within which we could explore the nature of the nonsense. To find some radical error in physics would be like finding that everyone had been talking nonsense - a discovery that could not be articulated, because there would be no language to articulate it with; a discovery that could not be scientifically demonstrated because the tools of demonstration themselves could not be relied on ...

In a way, our science says: 'If the world were a perfect machine which followed these (specified) rules, then it would behave in this way'. But the rules of this machine are explicit rules - rules which can be articulated. We cannot appeal, for their 'reliability', to some more metaphyscially fundamental machine, because the rules specifying this machine could not be written down in any language we could translate (and so in anything that we could recognise as a language at all). We can't make sense of the question 'what rules must I be following, in order to be able to follow a rule?', nor of the question 'what rules does the world follow so that we are able to securely describe it as following rules?'

Semantic games

The 'language game' metaphor has unfortunate overtones. It suggests triviality, and also that we might play the game in other ways - that we can make up rules any way we like. It also has another disadvantage, that is more directly related to Wittgenstein's exposition of it -

The 'pieces' in Wittgenstein's game - the words, sentences etc. - are caricatured as 'tokens', whose 'meaning' is given by their role in the game. This is fine, so long as we realise that this is a metaphor, and that it can' t be pushed too far.

If we're playing chess, it's easy to see how we might swap certain pieces - use the piece normally used as a knight as a bishop, and vice versa - and still be playing the same game (or one which is only trivially different). This is because we can usually identify the pieces separately from their role in the game. Someone completely unfamiliar with the rules of chess could learn the correct names for the pieces and be able to use them in our conversation (although in a specifically limited way). They could pick the right pieces from the box in response to appropriately worded requests, or they could make a list of 'all the chess pieces', or they could weigh them and come up with the same results as someone who could play chess etc.

This is not the case with language in general, even if it seems to be the case for some restricted games like chess (but see below).

We might think of a parenthesised or quoted word as a token - "word" or "knight" - for instance, and imagine a certain way of identifying it that doesn't take account of its 'meaning' (role in the game). This is necessary, for example, if we are to programme a computer to process strings which encode linguistic moves. My computer does not 'understand' what I am typing, but it must be able to reliably distinguish some strings of characters from others in order to function as an appropriate channel for our communication. The routines developed to achieve this can also be enhanced to do spell checking and even some translation.

Notoriously, we can also, to the point of practical incoherence, swap these pieces between roles without altering the 'meaning' of what we are saying.

What we cannot do is give an account of language use completely in terms of tokens and rules - even if we enhance this ('syntactical') account with a mechanical account of how certain tokens are related (in a functional, non-ambiguous way) to certain features of 'reality'. This is because (OQA) this account itself depends upon the correct 'intepretation' of the tokens which are used to encode it.

While we might demonstrate that some games (e.g. versions of chess in which token allocations are different) are 'isomorphic' in some important way, we can only give an account of this isomorphism, itself, within a playable language game. The syntactic/mechanical story says, roughly, that our whole language (and all intertranslatable with it) is isomorphic with some general set of syntactical rules and semantic/mechanical allocations and that it can be used to state these rules and define these allocations.

The 'game' is not like this - it is irreducibly semantic. If the rules of chess only defined playable moves in terms of the names of the pieces (without descriptions), the actual bits of hardware could look like anything at all, so long as they could be matched to the names - and the 'meanings' of the names would be given entirely in terms of the rules of play. In this circumstance, a non-player could not name the actual pieces - pieces in a box, and not on a board (in play), would not even have names, because the could be used to fill whatever 'logical' role we liked. And if a possible move in chess was to (re-)define how a piece could move ...

Irony and Gettier

The Gettier counter-examples depend upon someone's thinking that they have a justification for a belief when, unknown to them, they do not - even though the belief itself is true.

These depend upon an 'irony' in the sense that the audience for the counter-examples knows something that the subject in the counter-examples does not. Irony cannot survive audience participation, however - as demonstrated in pantomime.

The justification in a Gettier example cannot work in the context of a shared conversation between the subject and the audience - if I know that Farmer Franco has mistaken a large piece of black and white cardboard for his cow Daisy, then I cannot accept his justification that he has seen Daisy and so knows that Daisy is in the field. I can only accept that he believes it, and that his belief is true.

In any ('honest and competent') conversation with Franco, either what he saw, or what I saw, or (perhaps) what counts as justification, would have to be in play.

This is an exact parallel with the solution to Kripke's paradox, because it depends upon attribution of intentional states within a shared conversation.

Short Cruise

I sailed to St. Margaret's Hope and back betweeen 27/7 and 3/8 - here are some photos:

Orkney 2010

Everything worked well, but still looks pretty dreadful. A sailor from a nice Norwegian yacht in Wick suggested paint ... with the best intentions.

Rational attributions of intention

While no behavioural narrative can fully justify an attriubution of intention, it is clear that an attribution of intention can be inconsistent with a narrative. We may not have evidence that allows us to discriminate between adding and quadding, but still have evidence that woudl allow us to discriminate between adding and subtracting.

The possibility of the evidential narrative, itself, depends (of course) on some attribution of intention (to its narrator and to its audience). Maybe we can set this aside for the moment - allowing that some agreed behavioural/evidential narrative is possible, without enquring too much.

From this narrative, we can rule out certain possibilities (e.g. subtraction). Do we do this on the basis of 'facts', or only on the basis of some buried intentional attribution in the narrative?

If I say 'he wrote down this: "2+2=4"', I don't see how I can discard the meanings (e.g. by giving some description of the lines forming hte characters) without also losing the ability to rule out subtraction.

It might be the case that we can't describe behaviour without introducing (and so ruling out) some intentional content.

Although: a computer programme can look exactly like this. It may describe, in a complex way, how pixel elements on a screen can be made to change their character without saying 'this is how to display the letter "F"'. A slight peculiarity here is that the non-intentional explanation of what is happening is frighteningly complex - more complex than a human being could understand, if 'understand' means recognising what the programme is doing (printing "F").

In real programmes, we find a hierarchy of machines. At one very low level, there are machines for handling the pixels and changing their status. At a higher level, there are machines for assembling instructions to these low level machines which 'draw lines', or 'print characters'. Other machines help with re-directing these instructions so that we get 'print the character held in x at the present cursor position' - at which stage we are clearly dealing with instructions which have comforting intentional content. We may decide ourselves whether the linguistic structure here is just a convenient mnemonic, or whether we are succeeding in producing intentional behaviour on the part of the machine. This is a normative issue.

We can make a machine which 'appears' to talk to us, and we think we've achieved some insight into 'how talk works', but all we have done is make a machine behave in a way which invites an intentional interpretation by, among other things, appearing to talk and, particularly, appearing to talk in a way which rules out certain specific intentional interpretations.

Machine people

I can't calculate a paradoxical Gödel number. This is a 'fact' about me - about the kind of machine that I am?. We know, I think, that this is a fact about any possible machine. But, of course, he has proved that these numbers 'exist', despite this. They do not, of course, exist in any machine world - any world of of 'facts' - however organised. And so, in no 'possible' world of facts? And what would this mean?

I suppose, at least, that we shouldn't worry about them so far as 'practical' (machine manageable) arithmetic is concerned.

This machine world doesn't render up rules. It can't define them.

But a 'describable' or 'completely narrated' machine world's rules would be contained in its narration - we could only desribe it in this abstract way. Even a single determinate fact has a hidden rule - that a description of it is always true. If we can't narrate a world without some rules being true, then these rules might as well be in the world we are narrating.

But 'might as well be' ... this is metaphysics. It is 'as if' the world contains these rules. We can talk 'as if' the world contains these rules. Even: we could not talk except to talk as if the world contained these rules, except here the 'cannot' is a product of an argument within our talk, and is perhaps circular.

Talking machines (revisited)

Being able to talk is not the same as, say, being able to produce some string of characters (like the one which comprises the coding of this post). We can see two instances of the same string of characters and say of one 'this was produced by an interlocutor, and is part of a conversation' and of another 'this was produced by a machine, and is a recording or representation (etc.)'. In principal, it wouldn't matter how long this string was.

It couldn't be the coding for all human conversation to date, however, because it wouldn't mean anything if it did.

If we say 'the whole of human conversation to date is a string produced by a machine', we would have to wonder what language we were saying this in. If we are including it within its own scope - as part of human conversation to date - then either (a) it is false or (b) it doesn't mean anything because it's just a string of code. If we are not including it within its own scope, we might wonder how it, uniquely, could mean something if everything that looked as though it has meant something up to that point was just a string of code. A language game cannot comprise a single sentence standing on its own.

Should we think of ourselves as being the ('deluded') components of some greater machine which we cannot 'comprehend'? This could only ever be a metaphor - we could not describe the machine we were 'part of' well enough to call it a machine. Those elements of its workings we did not understand we could not distinguish from random, or at least radically unpredictable. And there would always be elements we did not understand because we are part of the machine, and we can't model ourselves. The machine cannot produce a string which encodes a complete description of itself (including its capacity to produce this string).

Who would this description be for? What is the language within which this 'representation' would work? (i.e. would be a legitimate move...)

A machine can't have a private language either.

The halting problem is also in here somewhere ...

Suppose that we imagine the machine to be following some rule - we see that the strings it produces, when interpreted in a certain way, do not breach the no contradictions rule. This wouldn't allow us to hand our adjudications on the rule over to the machine - it might not always behave correctly; it might break down. The machine can follow the rule, but it can't define it - and neither can any other machine.

(Maybe this is 'why' the halting problem arises ...)

What do we do when we say 'this is right' or 'this is wrong'? We state a rule that we are going to follow - we specify a part of the theory of truth we are using. Our justifications of these statements are always 'incomplete', in the sense that they cannot look outside our language, in which they are expressed. Our ability to agree empirical hinges is only different from our ability to agree logical or mathematical hinges in so far as it is more 'mysterious' to us (epistemologically). They are all rules which, if broken, render the conversation impossible - including, in extremis, any conversation in which we might be able to say why this conversation had become impossible.

Perfect Machines

A machine is an arrangement of facts into function, and whether 'facts' are intentional, funtion irreducibly is. The machine cannot define the function: 'or' is not 'what an or gate does', except in the sense that it's what it is designed to do. How do we know when a machine goes wrong?

This is just the 'naturalistic fallacy', slightly disguised ...

Another open question argument? But not quite as boring as that.

If I am a machine then 'or' cannot be defined in terms of anything I do either. What sensible conclusions can we draw from this?

There is the 'private language' issue - we know that 'or' can't be defined in terms of something internal - something I 'think'. But isn't a collection of machines also a machine?

This looks unavoidable, but what is the function of this 'greater' machine? We don't know. What would a 'function' of this kind look like? Function for what? To whose purpose?

We might say: a functionless collection of facts may still have some order - it may do something, but not something useful. What order would it have? Only the order of some facts. An ordered groups of facts is just another fact. We might make a mistake about the order, but this would not render the group of facts 'malfunctioning'. How would we discover this order? Just as we discover the ordering of facts which allows us to build our machines. The Universe doesn't get things 'right' or 'wrong'. The great collection of machines can't get things 'right' or 'wrong' either. It is just the way it is.

But we get things right and wrong. If we didn't, we couldn't talk to one another. Another dull chant?

Meta-epistemology

A scientific theory need only be presently intelligible - 'unfalsified', in traditional terminology. This can mean that, given other hypotheses, its contrary is unintelligible.

The contrary of an epistemological theory needs to be unintelligible tout court.

If it isn't, then the theory is hostage to any hypothesis on which its contrary's unintelligibility depends. More open questions ...

Tacit Rules

It's tempting to think of tacit rules as just rules we haven't stated - as though they are still in a cupboard somewhere, just not having been taken out and looked at yet.

This metaphor is misleading, though. Is it a tacit rule of chess that the normal rules stay the same at least until next Wednesday? Or next Thursday? We can ask an indefinte number of questions to which 'rules' like this might be answers.

The fact that 'normal' interlocutors know most of the answers to these questions (when asked) is like the fact that they give commensurate answers to questions about their immediate perceptions. They just know how to talk - or so we would say in any shared game with them.