Proving rules

There is a saying that 'the exception proves the rule'.  I used to find this very puzzling when I was young.

It means (I think) that an exception which honours the purpose of the rule helps us to understand how to follow it 'properly'.

In a Dworkinesque example, the rule 'do not walk on the grass' is made intelligible by the exception of the gardener's behaviour.  Because the gardener can walk on the grass, we understand the rule as having the purpose of preserving the lawn, and this helps us to follow it:  it makes it more intelligible even though (apparently) less universal - less mechanical.

If we want to construct a machine, however, we do not need intelligible rules, we need reliable rules.  We need to be able to use the rules to predict the behaviour of the machine.  Production line manufacturing, particularly in it's earliest manifestations, increased control and predictability, as well as efficiency (although efficiency dominated the PR).  A human production line worker is expected to behave in a machine-like way.

I expect that the tacit negotiations of just how machine-like would reflect important aspects of the power structures of particular institutions.

A  syntactical, mathematical, or 'physical' rule can look machine-like.  It operates over a restricted domain, and has explicit outcomes for all the possibilities within that domain.  It is 'embedded' in the description of the domain.  When we explain 'mass', we also explain 'acceleration' and 'force'.  We can write down some equations which exhaust these concepts and their interactions.

There are appropriate and inappropriate ways of applying this 'formalism' to the 'real world', of course, and the rules for doing this have to be intelligible rather than computationally reliable.

Compositional meanings, again.

Does a sentence have meaning?

Davidson seems to dispense with word meanings, but retain sentence meanings.  This is OK, if you think (as he does?) of words as affecting sentence meaning through composition rules, but not having meanings on their own.  But if we distinguish 'having a meaning' from 'having a role in generating meaning', we need to have some kind of criteria for doing this.

Clearly, for Davidson, the sentence/fact relationship won't work for this, because he accepts the relevance of the slingshot argument.

So why do sentences have meaning?  We have to be careful here:  if sentences only have meaning as individual elements in a playable language game, then they may be no better off than words or other sub-sentential elements.

But we can't think of the game as a whole as having 'a meaning'.  We can think of it as being meaningful.

Matrix questions and intelligibility

If the question 'Are we living in a matrix?' can be taken seriously, it cannot also be a question which it is only possible to ask in the matrix world.

If we think of the matrix world as a kind of deception, and that we are deceived that we are asking a question - including a question about whether we live in a matrix world - then we are not asking a question.  If we are 'really' asking a question, then it is a question which makes sense in the 'real' world and so is not, itself, a product of the matrix.

If we can ask the question, some part of us is cognitively independent of the matrix.

Verifiability ...

After hearing David Chalmers talking about The Matrix in Aberdeen:

This is probably not novel, but the verification theory of meaning had at least one thing to be said for it:  that a statement which cannot in principle be verified or falsified is meaningless.  The difference between the film matrix and the philosophical matrix (the brain in the vat) is that in the film Neo escapes from the matrix.  There is a narrative of discovery.

It's when we think we might live in a matrix from which we cannot escape, that we are thinking nonsense.  We are wondering, then, whether Neo 'really did' escape, or whether he only passed into another illusion - another manifestation of the 'real' matrix.

Without the narrative of escape - the possibility of making a test, at least in principle - the metaphysical speculation becomes meaningless.  With the possibility of a test, the 'metaphysical' speculation stops being metaphysical.  Perhaps we'd better not say it becomes 'scientific', as other tests might be appropriate.

There are some special cases - logical rules, for instance - where the alternative to their truth is that all statements become meaningless.  This prevents certain kinds of (experimental) speculation about alternatives to them, but still allows this argument for their truth to be made.  And in real cases where a potential interlocutor cannot be interpreted in a way which is consistent with their conforming to these rules, exactly their status as an interlocutor is called into question.

When we speak meaningfully, we are not claiming to have a fully worked out verification scheme in mind - far less a methodologically privileged one along the lines of the Logical Positivists' - but we also can't claim that no verification scheme could ever be available.  This claim is catastrophic for meaningfulness.

And, of course, both skeptics and religious believers do make this kind of claim:  the first that nothing can be proved and the others that certain things can never be disproved.

Open Questions

Open question paradoxes arise from questions that have this structure:

"What intelligible account can we give of the reliability of our intelligible accounts?"

What sensible story can we tell about how to tell sensible stories?

Formal systems within which questions with this character can be raised will be incomplete or inconsistent.  Even if we try to ask the question in a very disguised way - for instance by looking for a foundation for mathematics in set theory - our trick will be found out.

It cannot be sensible, or even intelligible, to say that we cannot give sensible or intelligible accounts - because our 'serious' statements must be accompanied by a tacit claim that they are, themselves, sensible and intelligible.  The claim that we cannot give intelligible account is either unintelligible or false.

So (a) we must be able to give intelligible accounts and (b) we cannot give a 'complete' account of why they are intelligible.

We can, however, found our accounts of the intelligibility of some statements on the intelligibility of others - even down to the intelligibility of logical axia, to the extent that these can be shown to be intelligible as a consequence of the necessity of intelligibility itself.

So that's OK, then.

However intolerable our private epistemic anxieties, we cannot intelligibly aritculate them.

Obvious arrangements

I have four socks which comprise two pairs.

Two of the socks have red toes, and two have green toes.

I bought these socks in a yacht chandlers.

Chinese Rooms, Translation, and Semantic Chaos

If I wanted to talk to you, and we spoke different languages, I might employ a translator - someone who spoke both your language and mine.

Translation software, of variable reliability, is also available.  We might look forward to this improving.

Suppose software existed which reliably translated your language into mine, and mine into yours.  Why would either of us, then, learn the other's language?  In fact, why would anyone learn both languages?

If no-one learned both languages, how would the software be maintained?  We could explore misunderstandings by discussing them with one another via the software ... which is approximately how we maintain our monoglot exchanges?

We feel as though we can ignore the problems this creates for settling the meanings of monoglot exchanges, because we do (a priori) manage to mean some definite things by what we say.  We always have the starting point of the present exchange that we are engaged in.

Would we be able to ignore these problems in a much more machine mediated system?  What about a world where every communication depended on mechanical mediation, where everyone needed a translation machine in order to speak at all?

It's not just a matter of hearing unfamiliar languages being translated here, but of having to treat what is going on as a language entirely on the evidence of the translation device.  And by 'what is going on' we can only mean 'whatever the device is translating' - which may be nothing we can independently discern ...

What could we say, in this environment, about the objects of the translation software?  There doesn't seem to be a coherent sense in which we could claim that it was preserving some kind of semantic content, since there would be no independent community of interlocutors to regulate this.  There would be no linguistic community that did not depend on the translation devices.

Phenomenology, again...

Hume, expressing a central Enlightenment norm, warns us against projecting our normative needs onto an indifferent reality.

Neither this thought, nor any other, can be articulated without observing the norms which make conversation possible; but these, themselves, can only be partly articulated, and provide only slight phenomenological comfort - so Hume's injunction has been practically reliable.  The moral certainties he was tilting at were misleading in proportion to the comfort they pretended to provide.

But we are left with (1) the necessity of conversational norms to philosophical theory and (2) the visceral intuitions to which we, as competent language users, are subject.  The possibility of articulating anything at all rules out certain skeptical positions, and we only respond to, rather than understand, the mechanism which drives our articulations.  We know we can talk, and we see the world as talkers see it.

Our phenomenological state, beyond supporting our conversations, fuels much wordless anxiety and confusion.  Perhaps we can take a perverse comfort from not being obliged to be able to account for every twinge.

Emergence?

If the biosphere is to include intentional creatures, then the problem of stating its boundary conditions is not to do with complexity or emergence, but is a consequence of our being part of the biosphere.

To be able to state boundary conditions is to be in a certain intentional state, and any statement of biosphere boundary conditions would, inter alia, be a statement of the boundary conditions for being in intentional states, and so of the boundary conditions for being in that one in particular.  This means that these boundary conditions would have to define what it was to be in an intentional state, and what it was to be in the intentional state of being able to state boundary conditions.  It would be a boundary condition and also define what a boundary condition was.

This can't work, of course, because such a boundary condition would be computationally incomplete, or self-referential.  We couldn't state the boundary condition without first stating a boundary condition for stating boundary conditions, and so on ...

Again, we have to give our fundamental accounts in terms which include 'primitive' intentional states which cannot be defined away.

Methods, Family Resemblances, and Rules of Reference

'Object Oriented Programming' is a structured programming approach adopted by software designers, and which is supported in fairly disciplined (although somewhat varying) ways by certain programming environments (e.g. Java, C++, Visual Basic, and others).

In these contexts, an 'object' can be thought of as a collection of data items, and a set of methods for accessing and manipulating them.  Each object has its own private set of data, and that data can only be accessed by calling on one of its methods - and not, for instance, through a public structure of variables or constants.

The power of this structure is that once an object, or rather an object 'class' - a type of object - has been constructed and tested, it can be relied upon in a large number of different contexts.  Because its methods and data are 'private' to it, their operations will not be interfered with in unexpected ways by other processes in the system, because the disciplines of the environment, or the structure of the software, prevent this.

What I want to say a little bit more about here is the concept of a 'method' borrowed from this environment.

Imagine (as in a classical example given in other expositions) that your objects are simulacra of vehicles in, say, a traffic network simulation.  Each of these vehicles will have methods associated with the kinds of things that vehicles do - we may want to feed them instructions about speeding up, slowing down, turning, stopping etc.  In interactions between them (collisions, collision avoidance etc.) we will want to define the ways that they will behave and calculate the outcomes.

For some instructions, the vehicles will react in different ways.  For instance, the instruction 'open the throttle 10%' initiates a method, with an argument, to a vehicle simulacrum.  The 'open the throttle' method of each vehicle type will tell it how to respond, and that response will be mediated by, and will alter, its internal data states.  A lorry, for example, will accelerate more slowly than a sports car.  A vehicle that already has its throttle fully open will not change its behaviour.  A vehicle with its engine switched off will not change its behaviour.  Some vehicles in certain icy conditions might lose adhesion.

Once we have constructed the classes of objects required (different types of vehicle) and instances of the classes (actual vehicles, with initial data states) the fact that all objects from certain classes have an 'open the throttle' method, simplifies the simulation.  We can, for instance, see what happens if everyone on the M6 between 7:30 and 8:30 in the morning were to open their throttles by 5%.

The relevance of this example is in the fact that there is no disciplinary or structural reason why methods with the same name should execute in commensurate ways over objects.  There is no reason why the method 'open the throttle' in some object classes would not produce bizarre behaviour (driving off into a field, deleting itself from the programme, sending a recording of  'Yankee Doodle' to the sound card ...).

Clearly the programme will be much easier to understand if different objects executed similar sounding methods in commensurate ways, but what counts as 'commensurate' for this purpose will be a feature of the programmer or of the reader and user, not of the objects or methods.  If these interlocutors find the ways that different object classes executed homonymal methods commensurate, then they are commensurate.

We might give a 'family resemblance' (after Wittgenstein) account of why a group of methods seemed commensurate.  This kind of semi-descriptive account can give the extension of a class, but at the expense of rendering membership decisions, themselves, in terms of methods which vary with the objects subject to the decision.  In other words, 'family resemblance' accounts really tell us no more about the way that methods are commensurate than the fact that they share a name does, since they, themselves, may depend upon a set of methods which we can only show to be commensurate by referring to some further family resemblance a little higher up the hierarchy of accounts.

Object oriented programming, then, has done something very odd.  In pursuit of reliability and predictability, it has created a place for natural language categories that it is hard to give a formal account of.  It has done this by allowing these to be coded in a variety of different ways, depending upon the objects they are to be applied to.  It has placed the decisions about how to organise these in an intelligible way in the hands of the programmers and users, so that this intelligibility is an aspect of the playability of the language game in which they are interlocutors, and not of structural features of the methods themselves.

It is sometimes useful to keep this model in mind when considering some of the 'hinges' of intelligibillity in general - hinges upon which anything that counts as a language must turn.

An example might be made, here, of reference and naming.  It's clear (although perhaps a more formal argument to this effect should be constructed) that it is necessary to be able to refer and to name in order to be able to speak at all.  Powerful instances of this suggest themselves - in particular the case of referring to interlocutors, but also cases of referring to and identifying the great range of  'objects' which are projected from nominal grammatical forms.  'Me', 'my pig', 'pig' and 'Piggy' (the name of my pig) all seem to need 'content'.

In the empiricist tradition, some of these contents have been sought in the world, and so the 'reference' relationship has seemed to have a special epistemological and linguistic significance.

If we think, instead, of 'reference' and 'naming' as being essential grammatical or semantic categories requiring contextual methods (as throttle manipulations might be to a traffic simulation), we can see that all this requires is that thes methods are adequately defined for the objects they are relevant to, and that this 'adequacy' is no more than that they are intelligible for, i.e. usable by, interlocutors.

It may still make sense to ask how reference works, or how naming works, but the answer to this question will be provided by an examination of the way the related methods are implemented with respect to specific linguistic 'objects', or computational contexts.  Any commonality would be provided by the grammatical and conversational needs, not by relationships between the methods (except in so far as these shed light on the conversational needs they might be meeting).

It is not surprising, perhaps, that specific 'methods' of naming (e.g. causal chain accounts) can seem very compelling in specific contexts.  References to actual historical figures must, after all, guarantee some narrative of this kind - so that where the narrative fails, the reference fails also.  References to fictional figures must, however, work in some quite different way, with different associated failures.  And there can be many differences of detail between methods - is Gilgamesh a fictional character in just the same way as are Zeus, or Mrs. Bennet, for instance?

Similarly, we can be tempted into a misleading simple mindedness about whether a proper name can function as a description - cases can vary, so long as the consequences of variation don't threaten local intelligibility.  We don't need to give a further account of how intelligibility in general is maintained, as this is guaranteed by our ability to request and give accounts.

This kind of mistake is a consequence of expecting some epistemological or metaphysical significance of a grammatical category, simply on the grounds that it 'works':  we think that because it works there must be a unifying underlying mechanism, when in fact grammatical necessity generates the requirement for mechanisms, for methods, which then may vary between classes of object/contexts.

A general demonstration that such methods must fail would be a demonstration that we could not render ourselves intelligible in a language which required reference, but this is not the same as demontrating that there must be some general method (however complex or conjunctive its definition).  It does have the consequence that we must be able to produce a method when one is needed, but that is a much weaker conclusion, and unlikely to have much fundamental significance.

An aim of the object oriented approach, after all, is computational robustness, not formal elegance.  The discovery that a method implementation is incoherent only affects the class of objects which incorporate that implementation.  Other classes sharing a 'commensurate' method (one which does something similar for those classes, and has a similar name) may not be affected at all.

References, and so referential failures, are accomplished in a variety of ways in order to achieve a grammatical object.  These 'methods' are defined in the contexts for which they are constructed, and exist to bring these contexts within the scope of a particular grammatical category.  It makes no more sense to ask how naming really works, outside of these contexts and this category, than it does to ask what 'open the throttle' really means outside the scope of the traffic simulator and the conversation of the programmers (and ourselves).

We can be confused by the grammatical necessity of reference within any language in which we might want to interrogate our theories and concepts (including the concept of reference) into thinking that we have found a special link between language and the world.  What we should see, instead, is the necessity of constructing links of this kind in some way or another in order to bring the world within the scope of our articulated interrogations.  (And this may be done tacitly, as well as explicitly.)

Obviously, an account of any one of these links cannot, itself, depend upon the link being accounted for; nor can a comprehensive dictionary of these accounts depend upon any subset of them for its intelligibility.  This means that the grammatical concept of reference, the only one that is 'necessary', cannot be accounted for in terms of the methods, even though it is only the methods which do the 'work' we require to be done, by setting out definite computational or narrative approaches to fulfilling the grammatical requirement.

Our story about the relationship between language and the world is not essentially different from other stories in language about the world.  It can't depend upon its own account of its possibility or integrity.

Intentionality, Function and Description

I don't think Dennett ever realy tells us how a bag of molecules can take an intentional stance, but his hierarchy has some appeal.

I offer the following version:

Physical
System (signalling)
System (function)
Intentional
Reflexive

And, as before, while we must attribute intentionality to an interlocutor, we may attribute otherwise if it 'works' (though always fallibly, from Kripke).

What I hadn't thought about properly until a talk with BT on Monday, is that there are circumstances where we must practically, as opposed to 'logically' (or to avoid undermining our own intelligibility), attribute intentionality (and function).  This is where lower level descriptions are just too complex.

Higher level biological systems are a good example:  a physical description of them woud be hopelessly complicated, and only a functional description is tractable.

With some computer systems, it is possible (depending upon the level selected) that even a system level descriptions would be too complex.  And it looks like brains might be like this too ....

These are still different reasons from the ones we have for treating interlocutors as intentional.  What I'm wondering, though, is whether there is some relationship here that I'm missing.

One of the reasons I'm wondering is that there isn't such a big difference between treating a system as law bound and treating it as functional.  Maybe.

And if we cannot maintain intelligibility without attributing intentionality on practical grounds, what then?

Indispensible Indexicals

(Maybe this is just more on Moore's Paradox ...)

Traditional approaches to formalising natural language have tended to treat indexicals as a kind of short-hand: as though they can always be substituted with non-contextual names or descriptions without changing the sense of the embedding expression.

I'm not sure whether this presumption has ever been thoroughly examined.  What would a language without indexicals look like?  If we always referred to ourselves by name, or by description, for instances ...

There would be an important epistemological difficulty, arising from the point I have made in earlier posts about Kripke's rule following argument:  While 'Mary is talking to John' is a corrigible intentional description of their behaviour, 'I am talking to you' is not.


Only the first and second person have this character - 'They are having a conversation' is subject to behavioural justification in a way in which 'We are having a conversation', or 'I am talking to you' are not.  We can probably alwasy replace 'they' with a context free name or description, but not so 'I' or 'you'.

If this is correct, then not only are the first and second person indexicals not dispensable, but their role is epistemologically fundamental.

Knowing the answer. Happy New Year.

If we could look into the future to see the answers to our questions, we would hear people uttering sublime truths without understanding.  They will even live their lives by them, and build and use machines which depend upon them.

But they will not have arrived at them via our questions, and, because they 'know' the answers, they will not understand our confusion.

We might live among them for a while thinking this, and then begin to wonder whether we had understood their language at all, or simply stumbled into a jungle of misleading homonymy.

Truth and Meaning

I think a puzzle for Davidson's approach is that we couldn't work out how to speak a language from a truth theory for the language on its own.  Since a truth theory for a language only allows us to distinguish true from false sentences of the language, it can only teach us what to assert and what to deny.  It is hard to see how context can get in here.

I don't mean just the kind of context that disambiguates indexicals, or gives sense to 'occasion sentences', but also the conversational context, the context that makes a certain utterance 'appropriate'.

Is there an important difference, here, between having a theory of truth for L and having a list of all the true sentences of L?  Isn't a theory of truth (as envisioned by Tarski and Davidson) just a way of generating such a list, or determining whether an arbitrary sentence is a member of it?

How would we make a transition from having such a list, or such a test, to being able to speak L?  We may not be able to give a full account of meaning in terms of use, but we can't give an account that denies its relevance.  If we don't know how to use an expression, we don't know what it means.

We can be confused by the fact that we might learn to speak French from a theory of truth for French written down in English, but we would oly be able to do this if we can, as well, transfer certain principles of appropriate use from the English context onto the French expressions that we have learned.

Even if these principles were, themselves, articulated in both English and in French, they would not be part of the hypothesised theory of truth for French.

A priori tautologies

On second thoughts, logically a priori statements are slightly interesting, because the rules that they ultimately depend on are only intelligible within practices which, themselves, are at least partly 'hinged' upon non-tautological a priori statements.

It is not possible to articulate a logic without a language to articulate the logic in, and the possibility of this language, if it is considered by its speakers, must be the subject of an a priori judgment.

The necessity of logical laws follows from the a priori possibility of the language in which they are entertained.

Distinctive definition of 'A Priori' ... tidying up

An 'a priori' statement is one which, though not a logical theorem, cannot be false in any playable language game.  Whether we call logical theora a priori is a matter of taste, but they are certainly not interestingly a priori.

Among necessary falsehoods, we can recognise hypotheses about the impossibility of language in general, claims which are Moorean 'paradoxical' etc.

There is room for us to argue about whether some particular statement is a priori or not - for "Neurath's Boat" type reasons, and because there is an open question bar to a general theory here.  Whether some statements are a priori may be a matter of linguistic experiment - of 'doing philosophy', perhaps.

Composition

Clearly, some 'concepts' are, at least partly, compositional.

We can't conclude from this that:

(a) we can recognise compositionality from particular structural features, or

(b) that the meaning of a composed 'concept' is exhausted by an account of its composition

This is partly because we can't sideline idiomatic uses.  'Brown suit' has a content which cannot be discovered from 'brown' and 'suit', because to wear a brown suit means something different than to wear a grey suit.  Or a pink suit.

It is possible that compositionality (ugh) is not even that important - that it sort of helps us along, but more as a kind of mnemonic than because we need it.

'But' the grammarians say 'how do you account for our indefinite capacity to construe meaning from a finite set of components?'.

The answer is in two parts:  what is the evidence that we have this indefinite capacity?  It doesn't exist, because an indefinite capacity can't be demonstrated.  It can only be a potential account - and an unlikely one, given our finite condition.

The second part is that even if we settle on requiring a rule-based account of human 'language' behaviour, this need not be a grammatical account.  A finite computational engine can produce speech-like behaviour from non-grammatical components - e.g. a list of sentences linked to a list of appropriate reponses.  Even were such an engine to do a little humanish parsing, this need not be its only method.

Dyslexic people sometimes have difficulty discriminating individual words - since they don't have the resource of a reliable capacity to decode or concoct written sentences, they are more dependent on phonetic elements.  For instance, they are more likely to think that common complex expressions - 'Just now', 'Good day', 'How are you' - are single words.

Most of us, even if not dyslexic, can remember similar errors from childhood.  I remember hearing 'Here we go round the' (... Mulberry bush) as 'Helligo rounda'.  The fact that it didn't have any clear meaning didn't especially distinguish it within my limited experience, and even adult literate reflection can't make a lot of relevant sense out of the first line of this nursery rhyme.

(Mulberries don't grow on bushes, for one thing.)

The compositionality argument has two parts:

(1) There are cases where we can deduce the meaning of an expression from grammatical rules and the 'meanings' of its parts (word meanings, in particular).

(2) We need some story about how we can decode expressions we haven't heard before.

But this is pretty ramshackle.  As above, (1) is rarely possible in a completely satisfactory way.  And (2) ignores lots of other resources we might draw on - particularly context, including any preceding linguistic exchange.

After all, that's how we start 'decoding' in the first place ...

Turing cognitivism

The halting problem has a structural similarity to the Goodman/Kripke paradox insofar as the fact that the machine has not yet halted cannot count as evidence that it will not halt.

Also, the demonstration that a human mind is like a general cognitive engine - a Turing machine - would have to include a demonstration that it was following the rules of such a machine, and Kripke has shown that this demonstration cannot be conclusively constructed.

So, curiously, it seems that if my mind is like a Turing machine, there is no finite way in which I could demonstrate this.  It would not be a computable issue.

The proposal that the mind is a kind of Turing machine can only ever be a kind of assumption of cognitive science - a 'false heuristic'? - and not part of a deductive cognitive theory.

'Meaning is Use'

We could confuse two views:

(1)  If we know how to use a word, then we know it's meaning.

(2) We give the meaning of a word by giving an account of its use.

The first seems unexceptionable.

The second is much more problematic.  To give an account, we are already depending upon some un-articulated meaning/uses.  Why should some of these be more 'obvious' than others?  Why, apart from a specific context of enquiry (a specific question, or specific ignorance) should some be based on others?

Also, we do not learn use meanings from such definitions - although these may aid us in some specific circumstances.  We learn them through practice, through experiments with use.

Also, approach (2) is confusing with respect to meaningfulness in general - knowing how to talk preceeds knowing that a certain articulated use of a word is correct.  It's perfectly easy to imagine a restricted language game in which the meaning relation didn't exist - in which no accounts of meanings were ever given.  As with the 'slab, brick' game, we might construct a theory of the meanings of the words in such a language.  But giving a type (2) theory in terms of uses rather than representations does not avoid the difficulty that Wittgenstein used this game to illustrate:  we cannot say that the users of the game share our theory of their meanings.

If we play chess with someone who cannot speak, can we say that they are following the rules of chess?  Kripke would say: only provisionally, and they might be following any number of other sets of rules which produced congruent behaviour in the games we had actually played with them.  Wittgenstein might say that they were not following rules, but just playing chess.  And that we might be wrong about that.

When we make a (1) type judgement - that we know how to use a certain word -  we are exposed to Kripke's scepticism about what we might mean by it unless we make it in a shared conversation which 'works'.  To ask whether this shared conversation might rest upon unrecognised Kripkean ambiguities is to ask whether it is, in fact, a shared conversation - a question which cannot arise within it.  When I say 'I know how to talk to you - I know what we mean by what we say', I am not saying something than can be refuted (shown to be false) by some behavioural evidence because a consequence of my being wrong would not be that I had said something false, but that I had not said anything meaningful - and so not anything at all.

So 'meaning is use' turns out to be a warning, not a theory - and, in particular, a warning against the semantic role of type (2) articulations.

Why should we expect unexamined heuristics to be false?

Here are some reasons why they might be:

(1)

Given Nature's indifference between functioning heuristics, and the likelihood that the ratio of successful false heuristics to successful true ones is likely ot be high, the probability that an unexamined heuristic is false must be high.

(2)

There is a sense in which all all language users share a specific false heuristic:  that they somehow live in the same world, and that this is why their shared language makes sense.  In fact, they only think they live in the same world because their shared language makes sense.  It is only in a shared language game that this question can arise, and it is only because the shared language game works that we can sustain the illusion of some pre-linguistic shared world.

(3)

Why would we articulate a tacit heuristic?  In normal conversation, this would only happen if a difficulty had arisen:  if interlocutors discovered that just pointing to a practice did not resolve a dissonance.  In this case, the interlocutors have discovered that they are unintelligible to one another, and need to modify the game to recover.

Except in the case of a simple grammatical or computational error, this process would have to result in at least some interlocutors abandoning heuristics as they articulated them.

(4)

The work of interrogating a heuristic falls outside the discipline the heuristic underpins - settling issues of its truth or falsehood is unlikly to be relevant to the discipline.  It might shed light on whether the discipline as a whole made sense, but this is not an issue routinely raised by practitioners, who can point to their own successful (?) practices in support of their intelligibility and who can therefore relegate heuristic 'housekeeping' to philosophers.  Their (reasonable) presumption might be that practically inconsequential inconsistencies in the underlying heuristics must be addressable.

An efficient intellectual division of labour would lead to disciplinary heuristics being simplified, since more generally defensible underpinnings would incorporate complexities which were practically irrelevant.