We all know that chatbots have come a long way in just a few years…
ChatGPT and other technologies like it are proving it to the world, using the power of LLMs to conquer bold new territories (or, perhaps more clearly, to boldly conquer new territories)…
But what can we expect from chatbots of the future?
We’ve talked about it a lot in places where we come together to discuss the potential of AI. To some extent, we’ve already seen big disruption around chat technology – but there are also all these question marks about how far it will go from here! You get that when you listen to dozens of entrepreneurs, researchers, and people connected to leading institutions give their pearls of wisdom to waiting crowds. And I’ve been doing a lot of that lately.
Regardless, what we’re seeing in terms of the evolution of chat is that many of these future chatbot systems will likely be connected to things that don’t look like large language models at all. Hmmm.
Let’s start with the basic principle of what these large language models do: they search a large amount of training data on the net, aggregate it, and use language as a tool to somehow mimic human cognition in digital environments .
In other words, they pass the Turing test to a high degree, due to a basic assumption we have about systems: we assume that if something writes something intelligent, it must BE intelligent .
Anna Ivanova addresses this topic by talking about chatbots and what we can expect in the future. It’s an interesting question, as Ivanova gets to the heart of how we interact with these intelligent chatterboxes. It addresses a “language network” in the brain with specific functionalities:
“It is a network that allows us to speak, to read, to write – and to understand others who are speaking,” she says.
She makes a very compelling argument: we assume that there are a lot of feelings and purposes behind what is written by a chatbot or generative AI simply because, as I mentioned above, we are used to assume that if someone makes a coherent statement, there is a lot of thinking about that statement going on behind the scenes.
But obviously this isn’t really true in the case of large language models. They aim, as Ivanova says, for coherence – but to the extent that they actually think, at least in the way we are used to, they have to recruit other systems.
Ivanova (who is a neuroscientist) and others also mention a “multiple-demand network” for handling complex processes, and it is suggested that if chatbots or AIs actually want to think like us, they will need to outsource things like an artificial system. limbic system and paleomammal cortex…
In other words, LLMs are great at language, but not so good at thinking, reasoning, or even math. The latter reason is partly because good mathematicians don’t think about their mathematics solely in terms of language. And LLMs always get all their answers from linguistic constructions they find on the Internet or elsewhere.
In other words, we are several steps ahead of a group of undergraduates who take thousands of images and ask the neural network to classify them, in order to know if something is a dog, a cat or not.
But LMMs sort of do the same thing with language: they make assumptions based on their training data, not the deeper purpose and intent tied to our fundamentally mammalian systems.
Although I say it often, Marvin Minsky said that the brain is not one computer, but several hundred computers linked together. We would do well to remember this as we enter the era of ever more powerful and capable AI. Where it will get really interesting is when scientists figure out how to connect an LLM-driven part to something equipped with our instinctual systems or a reasonable facsimile.