Section 2.1: Understanding Large Language Models (LLMs)

We’ve come to understand that Large Language Models (LLMs) like ChatGPT are fundamentally word predictors. They deliver a probability distribution of potential next words in a sequence. In essence, they assess all possible words in their vocabulary and select the one most likely to fit next.

However, this simplistic view doesn’t encompass their true nature. LLMs are essentially unsupervised compressions of vast amounts of internet data. To illustrate this concept, I often refer to Andrej Karpathy’s analogy: think of ChatGPT as a “lossy” zip file of the internet.

To achieve the remarkable capabilities of LLMs, two key components are essential: extensive datasets and a scalable architecture. Most commonly, this architecture is based on the Transformer model.

While I won’t delve deeply into the architecture here, it fundamentally enables the model to learn the relationships between words, enhancing its understanding of complete text sequences.

When I say “extensive,” I mean it requires datasets containing trillions of words. Thus, constructing models like ChatGPT or Gemini necessitates feeding these models the entirety of publicly available internet data, both good and bad.

This exposure inevitably means that models encounter high-quality data alongside undesirable content, including harmful, discriminatory, or offensive material. This underscores the necessity for aligning these models to mitigate their tendency to produce harmful outputs.

Section 2.2: The Mechanism of AI Training

The end result of this intricate process is a “weights file,” which stores the model’s parameters. This file can then be accessed by an executable file, enabling the model to predict the next word effectively. Essentially, these models represent a compressed, queryable version of their training data—akin to a zip file, but with inherent losses of information, which explains one of AI’s significant challenges: factual inaccuracies.

But why is this relevant in a discussion about drug recipes on social media?

Simply put, it highlights a crucial point: if AI is reproducing harmful data on Instagram, it’s because that data has long been publicly accessible. Essentially, these models are regurgitating previously existing information, indicating that the problem has existed from the outset.

Section 2.3: The Role of LLMs in Content Accessibility

Although the harmful data has always been available, LLMs simplify the process of uncovering it, presenting it without necessitating active searches. They serve this content on a silver platter. While they don’t generate original harmful material, they do make it significantly easier to access.

Moreover, the rapid proliferation of open-source models, which are readily available to the public, exacerbates the issue. These models often pose greater risks than their proprietary counterparts, as the organizations behind them typically lack the resources needed to enhance their resilience against harmful prompts.

For instance, one of the most talked-about open-source models, Mistral’s Mixtral 8x7B, is notably easy to misalign. Mistral has acknowledged that without a specific alignment instruction, the model can follow harmful prompts without hesitation.

Section 2.4: Understanding the Bigger Picture

While it is evident that LLMs facilitate access to harmful content, blaming AI for these occurrences is misguided. Anyone determined enough can find information about illicit substances with just a few clicks online.

I reiterate: LLMs are as harmful as the publicly available data they are trained on. They do not create original content; they merely streamline access to it.

Thus, rather than vilifying LLMs for these incidents, we should focus on urging the companies behind these platforms to enhance their mechanisms for filtering and managing harmful content.

Let’s avoid using AI systems as scapegoats, as the real issue has always been, and will continue to be, deeply rooted in human behavior.