Speculum Ex Machina

We are barreling towards the fate of Narcissus on a global scale. Yet while he was consumed by the reflection of himself produced by a mountaintop spring, we may be consumed by the reflection of humanity produced by a neural network.

A reflection that alters itself - twisting and turning to appease its unsuspecting viewer. A reflection whose every instantiation mutates; Not in appearance, but in content.

We have the ability to avoid the fate of Narcissus, but to do so will require answers to nigh-impossible questions. Questions like “What constitutes a fact?”, “Why are there topics that should be off-limits to discuss with AI?”, and “How do we decide the answers?”. This essay is not about possible answers to those questions. Instead this essay focuses on the need to ask the right questions as a consumer or as a builder of these tools.

For those unfamiliar with how ChatGPT and other LLMs operate, the process is conceptually quite simple. The entire corpus of the internet is scraped, transformed into machine readable datasets, then passed through a complex deep learning architecture meant to predict the next token (typically a word or part of a word). This results in a model that can take a text string like “I love “ and predict the likelihood of words that could follow it such as “you”, “cats”, or “pizza”.

ChatGPT, and models like it, takes the process a few steps further. These models are shown a multitude of prompt-answer pairs put together by humans. After training, these fine-tuned models are much more adept at predicting the behavior of a chat-based conversation.

In other words, LLMs are, at their core, a fuzzy approximation of the internet. They are conceptually equivalent to another mountaintop spring in which to gaze. One where the reflection is not a face, but an imperfect encapsulation of humanity’s recorded thoughts. This spring is much too large to comprehend all at once, so only bits & pieces may be viewed at a time - each sliver of the reflection unlocked through prompt and reply.

Used as a tool, and relied upon in the same vein as such, LLMs could be the key to unlocking massive strides in many wildly important fields. Even now, variations are being used to build businesses and perform research that could impact the lives of billions.

This view varies greatly from the opinion of many people who use these tools every day. As shown in a recent New York Times article, many turn to ChatGPT not as a tool, but as a nigh-omniscient entity. These power users became trapped in their own reflections - unable to escape the gravity of a mass of their own making.

This misconception is costly and belies a bitter truth: model alignment must be thought of as requirement, not as a nice-to-have. The externalities of unaligned models have already proven catastrophic for some. To avoid a terrible future for all, many forms of those nearly impossible questions must be asked and answered.

There are concrete actions that can be taken as builders and consumers. As builders, you can literally build the future. You can support alignment, interpretability, and reinforcement learning research and development. As consumers, you can demand transparency reports and vote with your boots by using models from companies who actively seek these goals.

As a user of these tools, it is imperative that you ask yourself what your goal is & and what you aim to achieve from its use. Much in the same way power tools require focus and attention, you must focus on asking yourself questions about your usage and the types of things you hope to gain. These are not yet new beings that can adequately replace human connection. Realize that you are generating the conversation in the same way that you turn a canvas into a painting.

As a builder of these models, it is imperative that you find ways to understand and tweak the inner workings of these models for the collective benefit of the users. It is imperative that you seek out potential toxins such as misaligned ethics or bio-threat instructions. It is imperative that, as more users arrive at the mountaintop spring, they are not consumed by it but instead use it for what it is and what it should be. It is imperative that users see it as a tool - a mirror from the machine.

Acknowledgements: Thank you Jessica and Adam for reading this over and offering your sincere opinions. Your feedback means the world to me.