LLMs aren't as good as humans at understanding, but it's not just statistics. The stochastic parrot meme is wrong. The networks create symbolic representations in training, with huge multidimensional correlations between patterns in the data, whether its temporal or semantic. The models "understand" concepts like emotions, text, physics, arbitrary social rules and phenomena, and anything else present in the data and context in the same fundamental way that humans do it. We're just better, with representations a few orders of magnitude higher resolution, much wider redundancy, and multi-million node parallelism with asynchronous operation that silicon can't quite match yet.
In some cases, AI is superhuman, and uses better constructs than humans are capable of, in other cases, it uses hacks and shortcuts in representations, mimics where it falls short, and in some cases fails entirely, and has a suite of failure modes that aren't anywhere in the human taxonomy of operation.
LLMs and AI aren't identical to human cognition, but there's a hell of a lot of overlap, and the stochastic parrot "ItS jUsT sTaTiStIcS!11!!" meme should be regarded as an embarrassing opinion to hold.
"Thinking" models that cycle context and systems of problem solving also don't do it the same way humans think, but overlap in some of the important pieces of how we operate. We are many orders of magnitude beyond old ALICE bots and MEgaHAL markov chains - you'd need computers the size of solar systems to run a markov chain equivalent to the effective equivalent 40B LLM, let alone one of the frontier models, and those performance gains are objectively within the domain of "intelligence." We're pushing the theory and practice of AI and ML squarely into the domain of architectures and behaviors that qualify biological intelligence, and the state of the art models clearly demonstrate their capabilities accordingly.
For any definition of understanding you care to lay down, there's significant overlap between the way human brains do it and the way LLMs do it. LLMs are specifically designed to model constructs from data, and to model the systems that produce the data they're trained on, and the data they model comes from humans and human processes.
You appear to be a proper alchemist, but you can't support an argument of understanding if there is no definition of understanding that isn't circular.
If you want to believe the friendly voice really understands you, we have a word for that, faith.
The skeptic sees the interactions with a chatbot as a statistical game that shows how uninteresting (e.g. predictable) humans and our stupid language are.
There are useful gimmicks coming out like natural language processing, for low risk applications, but this form of AI pseudoscience isn't going to survive, but it will take some time for research to catch up to understanding how to describe the falsehoods of contemporary AI toys
Understanding is the thing that happens when your neurons coalesce into a network of signaling and processing such that it empowers successful prediction of what happens next. This powers things like extrapolation, filling in missing parts of perceived patterns, temporal projection, and modeling hidden variables.
Understanding is the construction of a valid model. In biological brains, it's a vast parallelized network columns and neuron clusters in coordinated asynchronous operation, orchestrated to ingest millions of data points both internal and external, which result in a complex and sophisticated construct comprising the entirety of our subjective experience.
LLMs don't have the subjective experience module, explicitly. They're able to emulate the bits that are relevant to being good at predicting things, so it's possible that every individual token inference process produces a novel "flash" of subjective experience, but absent the explicit construct and a persistent and coherent self construct, it's not mapping the understanding to the larger context of its understanding of its self in the same way humans do it. The only place where the algorithmic qualities needed for subjective experience reside in LLMs is the test-time process slice, and because the weights themselves are unchanged in relation to any novel understanding which arises, there's no imprint left behind by the sensory stream (text, image, audio, etc.) Absent the imprint mechanism, there's no possibility to perpetuate the construct we think of as conscious experience, so for LLMs, there can never be more than individual flashes of subjectivity, and those would be limited to very low resolution correlations a degree or more of separation away from the direct experience of any sensory inputs, whereas in humans the streams are tightly coupled to processing, update in real-time, and persist through the lifetime of the mind.
The pieces being modeled are the ones that are useful. The utility of consciousness has been underexplored; it's possible that it might be useful in coordination and orchestration of the bits and pieces of "minds" that are needed to operate intelligently over arbitrarily long horizon planning, abstract generalization out of distribution, intuitive leaps between domains that only relate across multiple degrees of separation between abstract principles, and so on. It could be that consciousness will arise as an epiphenomenological outcome from the successful linking together of systems that solve the problems LLMs currently face, and the things which overcome the jagged capabilities differential are the things that make persons out of human minds.
It might also be possible to orchestrate and coordinate those capabilities without bringing a new mind along for the ride, which would be ideal. It's probably very important that we figure out what the case is, and not carelessly summon a tortured soul into existence.
LLMs aren't as good as humans at understanding, but it's not just statistics. The stochastic parrot meme is wrong. The networks create symbolic representations in training, with huge multidimensional correlations between patterns in the data, whether its temporal or semantic. The models "understand" concepts like emotions, text, physics, arbitrary social rules and phenomena, and anything else present in the data and context in the same fundamental way that humans do it. We're just better, with representations a few orders of magnitude higher resolution, much wider redundancy, and multi-million node parallelism with asynchronous operation that silicon can't quite match yet.
In some cases, AI is superhuman, and uses better constructs than humans are capable of, in other cases, it uses hacks and shortcuts in representations, mimics where it falls short, and in some cases fails entirely, and has a suite of failure modes that aren't anywhere in the human taxonomy of operation.
LLMs and AI aren't identical to human cognition, but there's a hell of a lot of overlap, and the stochastic parrot "ItS jUsT sTaTiStIcS!11!!" meme should be regarded as an embarrassing opinion to hold.
"Thinking" models that cycle context and systems of problem solving also don't do it the same way humans think, but overlap in some of the important pieces of how we operate. We are many orders of magnitude beyond old ALICE bots and MEgaHAL markov chains - you'd need computers the size of solar systems to run a markov chain equivalent to the effective equivalent 40B LLM, let alone one of the frontier models, and those performance gains are objectively within the domain of "intelligence." We're pushing the theory and practice of AI and ML squarely into the domain of architectures and behaviors that qualify biological intelligence, and the state of the art models clearly demonstrate their capabilities accordingly.
For any definition of understanding you care to lay down, there's significant overlap between the way human brains do it and the way LLMs do it. LLMs are specifically designed to model constructs from data, and to model the systems that produce the data they're trained on, and the data they model comes from humans and human processes.