Large language models (LLMs), a.k.a. AI, can produce plausible discourse based upon statistical properties of very large quantities of discourse produced by humans. Often, what they ‘say’ accords with our experienced reality because (or to the extent that) the input discourses are in this sense ‘factual’, but even if all the input texts were empirically accurate these LMMs can ‘overgenerate’ falsehoods and ‘hallucinations’. LLMs have been generalized from language data to data of many other kinds, notably visual (pixels) and auditory (sound spectra).
Such data correspond to but are not identical with the intensity and sensation perceptions constructed at the base of the perceptual hierarchies of biological nervous systems. The statistical constructs within an LLM are not open to inspection; they can only be inferred indirectly. Nevertheless, we may say with some assurance that the ‘reinforcement’ processes in ‘training’ LLMs, and the resulting statistical constructs in LLMs, correspond at least superficially, and possibly in fact, with how reorganization establishes perceptual input and output functions and reference values in living control systems.
In living things, however, perceptual input interactively with the environment provides a reality check on results of probabilistic reorganization. (If you don’t think it’s probabilistic, you haven’t investigated how synapses and neural brachiation work.)
Here is some recent work addressing this fundamental deficiency by embodying a very simple LLM in a robotic device that interactively perceives aspects of its environment.
A less technical discussion:
LLMs are now used to train other LLMs, and the corporations developing LLMs routinely pirate the accomplishments of their rivals by using their available products to help train their own.
Some background on DeepSeek and other recent developments:
