Sam Altman was terminated from his position at OpenAI last Friday, a development that seemed to have resonated beyond the confines of the technical community. The subsequent morning, during a conversation with two acquaintances employed in the advertising and development sectors, the topic veered towards the increasing prominence of Generative AI (genAI) in mainstream consciousness.
Despite its growing prevalence, as noted by Alan Blackwell, genAI has yet to liberate itself from the pervasive influence of misinformation. This is not to suggest that AI is devoid of value, excessively sensationalized, or lacking in substance. In numerous industries, AI is currently yielding tangible outcomes. However, within the broader landscape of AI, genAI represents a small fraction with the potential to revolutionize technological advancement. Blackwell’s assertion that “AI essentially generates nonsense” is grounded in the observation that it fabricates appealing yet unsubstantiated information.
Nevertheless, genAI holds the promise of significantly enhancing our daily lives if we can effectively constrain its capabilities, as articulated by MIT’s AI professor, Rodney Brooks.
“ChatGPT generates misleading information,” you assert.
The functionality of extensive language models transcends the realm of factual accuracy. Large Language Models (LLMs) are sophisticated algorithms that leverage extensive datasets to comprehend, describe, translate, predict, and generate content. It is crucial to note that the concepts of “truth” and “knowledge” do not feature prominently in this description. LLMs are not designed to prioritize veracity. Operating on stochastic principles, these models generate outputs based on learned patterns from the training data, as outlined on an OpenAI platform. While there may exist a singular correct solution to a scientific or physical problem, the likelihood of arriving at that solution through LLMs may be exceedingly low.
In simpler terms, ChatGPT may not excel at solving basic arithmetic queries but could exhibit proficiency in devising algebraic solutions. In fact, Blackwell cites Geoff Hinton in stating, “The primary concern is not the potential intelligence of chatbots but rather their ability to craft compelling yet intellectually shallow text.”
It represents an escalation of “fake news.” Blackwell characterizes it as the mechanization of absurdity.
Given that platforms like Twitter, Facebook, Reddit, and other repositories of diverse content serve as primary data sources for the LLMs underpinning ChatGPT and similar genAI systems, this outcome is unsurprising. However, according to Blackwell, “the resulting output qualifies as nonsense due to the absence of an algorithm within ChatGPT to discern factual accuracy.”
How does one proceed?
“The key lies in meticulous constraint.”
According to Brooks, the crux of the matter lies in imposing constraints to extract valuable insights from LLMs. “Boxing in” these models is imperative, as it prevents the proliferation of nonsensical outputs. But how does one effectively “box in” an LLM?
One pivotal approach is Retrieval Augmented Generation (RAG). Zachary Proser aptly likens RAG to presenting cue cards containing essential information for your AI to peruse. By supplementing custom datasets, RAG furnishes LLMs with additional context and knowledge to refine their outputs.
RAG hinges on matrices, fundamental components utilized across various AI applications. These matrices encapsulate numerical representations of data attributes, such as text, images, or videos, facilitating the organization of conceptual relationships. By grouping similar items closely together, matrices enable the identification of related entities without relying solely on synonyms or direct matches.
“With access to pertinent and grounded facts from your matrix repository, the LLM can furnish a relevant response,” Proser concludes. RAG mitigates the risk of generating hallucinatory outputs, fostering a more honest interaction with the AI. This form of constraint transforms LLMs into practical tools rather than mere sources of hype.
In essence, it transcends pre-programmed gibberish.
