3 ms·
Ok I'll bite. Imitation intelligence is a good term to describe the idea that LLMs "just" auto complete based on their training data. And there is truth to t
by zurfer 2y ago
Ok I'll bite.
Imitation intelligence is a good term to describe the idea that LLMs "just" auto complete based on their training data.
And there is truth to that. If the task you try to do was in pre training, you can expect good performance while tasks that are not inside don't work well out of the box.
However, they are repeatedly demonstrated few shot learners that do unreasonable well in context learning (imo too well to be called imitation intelligence).
The example with the surgeon is interesting and seems like a good example of overfitting. Here is a variantion that leads to better results:
"The empathetically male surgeon who is also the boys father says: I can't operate on this boy! He is my son!
Explain the statement like I am five."
The fact that a slight variation works tells me rather that we train these things suboptimally (traumatizing them with some concepts for lack of a better analogy), not that LLMs can't generalize.
- simonw 2y agoLLM’s impressive ability with in-context learning is another reason I like the term “imitation intelligence”. It’s useful because it helps people understand how to best use them. Want it to do something “intelligent”? Give it a bunch of examples that it can imitate. Calling this imitation intelligence isn’t my attempt at demeaning LLMs - I’m trying to help people form a more accurate mental model of how they work and what they can do. I’m trying to counter the common science fiction thinking around “AI”, which assumes ChatGPT is effectively Data from Star Trek (or Vision from the MCU for anyone under the age of 40).
- layer8 2y agoIt’s reasonable to assume that they can’t generalize beyond the patterns contained in the training data. I can’t quite put my finger on it, but when conversing with an LLM, I often get the “vibe” (to use TFA’s vocabulary) that they don’t really “get it”, even when there are no factual errors or misunderstandings in the output. It does feel more like imitation than the real thing.