Path: utzoo!utgpu!jarvis.csri.toronto.edu!mailrus!csd4.milw.wisc.edu!lll-winken!uunet!mcvax!ukc!strath-cs!glasgow!gilbert From: gilbert@cs.glasgow.ac.uk (Gilbert Cockton) Newsgroups: comp.ai Subject: Re: Where might CR understanding come from (if it exists) Message-ID: <2705@crete.cs.glasgow.ac.uk> Date: 31 Mar 89 10:38:17 GMT References: <2691@crete.cs.glasgow.ac.uk> <813@htsa.uucp> Reply-To: gilbert@cs.glasgow.ac.uk (Gilbert Cockton) Organization: Comp Sci, Glasgow Univ, Scotland Lines: 123 >This discussion has degraded into a fight between two groups with different >viewpoints: There is considerable diversity, as well as incompatability, in the arguments both for and against the possibility of strong AI. You are particularly poor in your grasp of all the anti-AI arguments. Some are based on the impossibility of simulating the brain's hardware on a digital computer (indeed, the impossibility of accurately and faithfully simulating ANY part of the natural world on a computer). Others rely on epistemic arguments. Others rely on theories of ideology which deny any possible objective status to such value laden concepts as 'intelligence', which are symptomatic of a system of social stratification peculiar to modern Europe (and taken on in an even cruder form in the New World). The word doesn't have a usable meaning in any scientific context. The sensible approach is to identify tasks for automation, describe them accurately and acceptably, and then proceed. Designing systems to possess ill-defined and hardly understood properties is an act of intellectual dishonesty at worst, and an act of intellectual mediocrity at best. Robotics has the advantage of dealing with fairly well-defined and understood tasks. I'd be surprised if anyone in robotics really cares if their robots are 'intelligent'. Succesful performance at a task is what matters. This is NOT THE SAME as intelligent behaviour, as we can have clear conditions for success for a task, but not for intelligent behaviour. Without verification or falisification criteria, the activity is just a load of mucking about - a headless chicken paradigm of enlightenment. >with "survival of the fittest", you have to behave in such a way that you >will survive long enough to raise a new generation. As the level of >complexity of the organism increases, it will have to do more "information >processing": to find food, to protect against enemies, etc. My point: >intelligence etc. developed out of a need to determine how to behave in >order to survive. So the behaviourist approach is justified: "when the >system seems to act intelligently, it *is* intelligent". You equate intelligence with a high degree of information processing (by co-location of sentences, there is no explicit or clear argument in this paragraph). A cheque clearing system does a high degree of information processing. It must be intelligent then - and AI was achieved 20 years ago? You are making a historical point. Please make it like a competent historian. Otherwise leave evolutionary arguments alone, as you are just making things up. >Before we can write a program for this, we must >understand the algorithm humans use. This proves to be very difficult. >Research is hindered by people claiming that understanding requires very >mysterious causal powers which computers, due to their design, can never have. 'Mysterious' is true only in the sense that we do not yet understand them. 'Eternally mysterious' would not be true. What is true is that causation in human/animal behaviour, and causation in physics, are very different types of cause (explanatory dualism). This does not hold up research at all, it just directs research into different directions. Logical necessity is a further type of pseudo-causation. Its relation to human agency is highly tenuous, and it is wrong to bet too much on it in any research into psychology. Computers cannot uncover mysteries. Automation research may do, in that the task or problem must be properly studied, and it is this study, which advances knowledge rather than the introverted computer simulation. Attempts at computer simulation do, however, expose gaps in knowledge, but this does not make the mystery go away - it only deepens it. The problem is that, if studies are driven by the imperative to automate, this will force the research into an epistemic and methodological straightjacket. This is a narrow approach to the study of human behaviour, and is bound to produce nonsense unless it balances itself with other work. Hence AI texts are far less 'liberal' than psychology ones - the latter consider opposing theories and paradigms. >Gilbert Cockton even claims that because human minds are not >artifacts, while computer systems always will be, there will always be >performance differences. Apart from the fact that this statement is >nonsense, it is not of any importance to AI-research. It is highly relevant. I take it that you think it is nonsense because I offer no support (reasonable) and you don't want to believe it (typical). An artefact is designed for a given purpose. As far as the purpose is concerned, it must be fully understood. The human 'mind' (whatever that is - brain? consciousness? culture? civilisation? knowledge?) was not 'designed' for a given purpose as far as I can see (i.e. I am not a convinced creationist, although nor have I enough evidence to doubt some form of creation). As 'mind' was not designed, and not by us more importantly, it is not fully understood for any of its activities ('brains' are of course, e.g. sleep regulation). Hence we cannot yet build an equivalent artefact until we understand it. Building in itself does not produce understanding. I can expose ignorance, but this is not cured by further building, but by further study. Strong AI does not do this study. My argument is initially from scepticism. I extend the argument to all forms of (pseudo-)intellectual activity which cannot improve our understanding. Strong AI, as modelling without study, i.e. without directed attempts to fill gaps in knowledge by proper, liberal, study, is one such dead-end. Computer modelling based on proper liberal study, is more profitable, but only as a generator of new hypotheses. It does not establish the truth of anything. Finally, establishing the truth of anything concerning human agency, is far far harder than establishing the truth about the physical world, and this is hard enough and getting harder since Quantum interpretations. We have insitutionalised research. There are areas to be studied, and a permanent role in our societies for people who are drawn to advancing knowledge. Unfortunately, too many (of the weaker?) researchers today see any argument on methodological grounds as an attack on research, an attack on their freedom, a threat to the advance of scientific knowledge, a threat to their next funding. The purpose of research is to advance knowledge. Advancing knowledge requires an understanding of what can, and cannot, count as knowledge. In our bloated academia, respect for such standards is diminishing. Research is not hindered by ideas, but by people acting on them. If strong AI cannot win the arguments in research politics, then tough, well - ironic really, for without research politics, it would not have grown as it did in the first place. Those that live by the flam, die by the flam. -- Gilbert Cockton, Department of Computing Science, The University, Glasgow gilbert@uk.ac.glasgow.cs !ukc!glasgow!gilbert