The robots are getting stressed out

The idea of machines expressing human traits is fascinating, and I think the challenges around this make some of the most compelling characters, at the very least in science fiction. Perhaps there is something inherently relatable about the challenges humanity faces that we recognize in robots and computer models when they begin to encounter and overcome similar obstacles. As these technologies continue to progress, I suspect we'll push the boundaries of what we know about ourselves even more than what we learn about designing computer models.

I've seen a few articles that cite a recent study on how large language models (specifically GTP-4 in the case of this study) begin to behave differently when presented with emotionally-charged, or stress inducing information. Specifically they become less likely to behave objectively and trend towards expressing biases latent in the data they've been trained with.

One suggested explanation is that this is in-part a "trait" of the LLMs. They've been trained using vast amounts of human generated data, and so they mimic the pattern that humans follow in similar situations. Somewhat humorously, the study showed evidence that using mindfulness meditation techniques on these LLMs helped to "relax" them and counteracted the trend towards biased behavior.

While the study focused on more of a black-box method of working with the GTP-4 model using prompts, I think there's something inherently interesting about the nature of the type problems that were presented to these LLMs.

Recently, I've been working on adding support for vector databases to the chatterbot Python library, and so naturally I've been giving a lot of thought to the idea of vector representations of information. But what are vectors used for in LLMs? In a general sense, "words" as the LLM understands them are represented as vectors, or rather a number that represents that word's relationship to other known words. As a small example, we can consider the relationship between "speaking" and "writing" alongside actions that could be synonymous with either:

        (Speaking)
            ●
           / \
          /   \
(Poetry) ●-----● (Rhyming)
          \   /
           \ /
            ●
        (Writing)

Speaking and writing have very close contextual relationships but on the other hand something like "running" or "battling" would be a significant distance away from these terms. 

Vector relationships like these may be relevant to the responses of LLMs becoming less optimal when presented with stressful information when we start to think about the attributes of those types of situations. Our own brains excel when faced with information and scenarios that we're familiar with, but begin to experience stress when faced with fundamentally challenging data. Much like the vector spaces, stressful scenarios involve points of information that are further apart, they are challenging because it takes greater effort to try to reconcile it issue at hand.

The distance between vectors is still a bit speculative on my part, but might be worth some experimentation to better quantify. More likely than not, the landscape of information shared between humans and machines is highly nuanced, and shaped by an extensive range of both strong and subtle forces.

No comments:

Post a Comment