Why GPT-3 Can’t Understand Anything

There is a mathematical reason why machine learning systems like GPT-3 are incapable of understanding. The reason comes down to the fact that machine learning has no memory. It is just probabilistic associations. If there is only a 10% chance of going off topic, then after just seven exchanges there is a greater than 50% chance the machine learning model has gone off topic. The problem is that when prediction is just based on probabilities, the likelihood of making a misprediction increases exponentially. A long-term memory is needed in order to maintain long-term coherence.

GPT-3 is essentially a sophisticated Markov process . What is important about the Markov process is that the next step in the process is only dependent on the immediate previous step, or a fixed number of previous steps. There is no longer-term memory of the past that shapes the future.
On the other hand, the distinctive characteristic of understanding is long-term memory. When two people are talking about a topic, and understand each other, then the present conversation is on the same topic that it was in the past, regardless of how far in the past the conversation started.

This means that the probabilistic models, like GPT-3, are incapable of understanding because they are inherently incapable of long-term memory.

This is also a surefire way to identify that a human is in the loop. If a chatbot is able to maintain a long-term conversation and consistently provide coherent responses relevant to the subject matter, then the chatbot is human-powered.

I’ve given a couple examples (see here and here ) of GPT-3’s ability to provide coherent responses over a long conversation, and that this is possible because humans are on the backend. As a contrast, I also conducted a couple conversations with GPT-J, which is the open source alternative offered by EleutherAI. You can try it out on the playground here .

It is a night-and-day difference. The GPT-J can’t put together a single coherent response. The closest it gets is when there is a text snippet in its vast lookup table that is associated with a question. The only way GPT-J provides the semblance of understanding is by relying on canned human intelligence, which again vindicates my claim that only humans generate understanding.

So there you have it. GPT machine-learning algorithms and the like are fundamentally incapable of understanding due to the way they are constructed from probabilistic associations. Statistics can only regurgitate existing human understanding in the dataset, and otherwise goes off the rails exponentially quickly. Machine learning is not up to the task of ushering in the singularity. And, by the way, this means the human mind is not a neural network.