Stevan Harnad

Apart from what (little) OpenAI may be concealing from us, we all know (roughly) how ChatGPT works (its huge text database, its statistics, its vector representations, and their huge number of parameters, its next-word training, and so on). But none of us can say (hand on heart) that we are not surprised by what ChatGPT has proved to be able to do with these resources. This has even driven some of us to conclude that ChatGPT actually understands. It is not true that it understands. But it is also not true that we understand how it can do what it can do. I will suggest some hunches about benign biases: convergent constraints that emerge at LLM scale that may be helping ChatGPT do so much better than we would have expected. These biases are inherent in the nature of language itself, at LLM scale, and they are closely linked to what it is that ChatGPT lacks, which is direct sensorimotor grounding to connect its words to their referents and its propositions to their meanings. These convergent biases are related to (1) the parasitism of indirect verbal grounding on direct sensorimotor grounding, (2) the circularity of verbal definition, (3) the mirroring of language production and comprehension, (4) iconicity in propositions at LLM scale, (5) computational counterparts of human categorical perception in category learning by neural nets, and perhaps also (6) a conjecture by Chomsky about the laws of thought. The exposition will be in the form of a dialogue with ChatGPT-4.

How do internal symbols become connected to the object they stand for?$\sp1$ A symbol system is a set of physical objects or states and the formal rules for manipulating them. The rules are syntactic, operating only on the shapes of the symbols, not their meanings. Yet the symbol combinations can be given a systematic interpretation or states of affairs ). These meanings, however, are not "grounded"; they derive from the mind of the interpreter of the symbols. How can the meanings of symbols be grounded without the mediation of an interpreter? This is the "symbol grounding problem." ;Symbol grounding is not possible in a pure symbol system, but in a hybrid system with nonsymbolic robotic capacities such as transduction, analog transformation and sensory invariance extraction, the meanings of elementary symbols can be grounded in the system's capacity to discriminate and categorize the analog objects and states of affairs that its internal symbols stand for, based on the invariants in the sensory projections of those objects and states of affairs. The grounded elementary symbols can then be rulefully combined to form higher-order symbol combinations that inherit the grounding as nonarbitrary constraints on their shapes. ;"Categorical Perception" occurs when there is a special interaction between discrimination and identification performance : Stimuli that are in the same category and have the same name resemble one another more than stimuli that are in different categories and have different names, even when the physical differences between them are equal. This "warping" of similarity space, the hallmark of CP, suggests a model for grounding higher-order categories in sensory categories: In a three-level representational system the top, symbolic level, can be grounded bottom-up in the lower two, the analog and categorical levels. ;This model was tested with neural net simulations using one-dimensional inputs: When a net is trained to discriminate and then categorize a set of stimuli, the categorization task is accomplished by "warping" the similarity space , the same natural side-effect found in human and animal CP. These compressed, identified regions of similarity space may be the ground level out of which higher-order categories are constructed. Nets are one possible mechanism for learning the sensorimotor invariants that connect arbitrary names to the nonarbitrary shapes of the objects they stand for, thereby mediating the constraints exerted by the analog world of objects on the formal world of symbols. ftn$\sp1$This thesis is based on four published papers

Steels & Belpaeme's (S&B's) simulations contain all the right components, but they are put together wrongly. Color categories are unrepresentative of categories in general and language is not merely naming. Language evolved because it provided a powerful new way to acquire categories (through instruction, rather than just the old way of other species, through trial-and-error experience). It did not evolve so that multiple agents looking at the same objects could let one another know which of the objects they had in mind, co-coining names for them on the fly.

Explaining the mind by building machines with minds runs into the other-minds problem: How can we tell whether any body other than our own has a mind when the only way to know is by being the other body? In practice we all use some form of Turing Test: If it can do everything a body with a mind can do such that we can't tell them apart, we have no basis for doubting it has a mind. But what is "everything" a body with a mind can do? Turing's original "pen-pal" version (the TT) only tested linguistic capacity, but Searle has shown that a mindless symbol-manipulator could pass the TT undetected. The Total Turing Test (TTT) calls for all of our linguistic and robotic capacities; immune to Searle's argument, it suggests how to ground a symbol manipulating system in the capacity to pick out the objects its symbols refer to. No Turing Test, however, can guarantee that a body has a mind. Worse, nothing in the explanation of its successful performance requires a model to have a mind at all. Minds are hence very different from the unobservables of physics (e.g., superstrings); and Turing Testing, though essential for machine-modeling the mind, can really only yield an explanation of the body.

The usual way to try to ground knowing according to contemporary theory of knowledge is: We know something if (1) it’s true, (2) we believe it, and (3) we believe it for the “right” reasons. Floridi proposes a better way. His grounding is based partly on probability theory, and partly on a question/answer network of verbal and behavioural interactions evolving in time. This is rather like modeling the data-exchange between a data-seeker who needs to know which button to press on a food-dispenser and a data-knower who already knows the correct number. The success criterion, hence the grounding, is whether the seeker’s probability of lunch is indeed increasing (hence uncertainty is decreasing) as a result of the interaction. Floridi also suggests that his philosophy of information casts some light on the problem of consciousness. I’m not so sure.