Context: Quick thoughts on GPT3
An aptitude to compose text artefacts without enduring, substantive (or perhaps any) comprehension forms the basis of many of our journeys through educational institutions. In a sense – blind rule-shuffling in context is what is requested of many if not most students in any standardised learning environment. Competency is measured on contextual cues and on acquiring a superficial appearance of possessing depth or substance of comprehension.
The language-as-statistical-artefact turn is quite insightful but not necessarily for the reasons we think. Are we moving closer to developing an emulated subjective (i.e. General) agency via functional inference? Is human cognition, communication and culture itself really only the aggregate consequence of patterns of complex probabilistic emergence and is this a hollow truth we are inexorably hurtling towards?
Complex information systems autonomously encode optimal heuristics for self-replication, in many cases via ambiguity and uncertainty. There is an argument to be made that whatever intelligence there is in GPT3 is a property, function and probabilistic potential in and of the languages themselves, not of the (synthetic or organic) agents/instances or the architecture that wields them.