Is it more appropriate to think of neural network based AIs as “Artificial Instinct” rather than “Artificial Intelligence?”#

I realized recently that a much better fit than intelligence might be instinct. The prompt in this analogy is like a novel external stimulus–a shadow, pangs of hunger, pain. These sensory stimuli trigger neurons to fire which results in signals to run, to seek food, to recoil. These responses are not learned. They are innate. There is no thinking, no analysis. There is just reacting. This seems like a better analog of what LLMs are doing. I find it difficult to conceive of what ChatGPT, Gemini, Claude, etc…, are doing as thinking and therefore as intelligence,

A large language model is presented with a prompt. This prompt is converted into tokens. These tokens are turned into a vector which activates the first layer of the neural network. This initiates a series of vector multiplications, which results in a value that is mapped to an output token. The process repeats until some exit criteria is reached. I have always struggled to reconcile this with what I think of as intelligence. When I say ‘always’ I mean since I first learned about how neural networks work. That’s not to say that I have some rigorous definition of what intelligence is, but whatever that might be does not fit whatever LLMs appear to be doing.

I have often wondered how newly-born creatures seem to ‘know’ how to do anything. Or, how creatures raised away from their natural habitat can seem oddly at home when repatriated. Spending some time looking into how large language models and neural networks operate has helped me conceptualize how instinct might operate. Over millions of iterations, evolution shaped the brain into a structure that when activated in a specific way results in a beneficial response to the host. Over untold iterations, brains that respond correctly survive to reproduce and others are eliminated. The offspring of the survivors inherit the beneficial responses. When that offspring is triggered by a similar stimulus it will respond in a manner that helps its survival.

Model pre-training and training take the place of evolution and natural selection. During LLM training, incorrect responses result in errors that adjusts model weights. This could be thought of as a specimen perishing and not reproducing. When a training input results in the expected output, that is the equivalent of a surviving offspring. The result is a model (neural network, brain) with a structure (layers, weights and, biases) that is able to respond correctly to some stimulus. In the case of a brain, that stimulus could be some event in its environment. In the case of an LLM that stimulus is a prompt.

I am not a computer scientist, biologist, philosopher, or neuroscientist and all of that is speculative and subject to correction. If I am anywhere in the ballpark, what might the implications be to achieving AGI? I’m not sure that the leap from instinct to intelligence is merely a matter of scaling or number of data centers.