Over the past year, large language models (LLMs), despite their decade-long existence, have captured significant attention for good reason. Cutting-edge models like ChatGPT 4 have emerged with advanced language capabilities, enabling more natural and contextually aware text generation. These LLMs can now produce highly realistic and diverse textual content, blurring the distinction between human- and computer-generated text. It’s not an exaggeration to state that generative AI is revolutionizing our interactions with technology. However, realizing the full potential of AI necessitates a comprehensive understanding of its characteristics and limitations.
The intrinsic nature of LLMs results in non-deterministic outputs. Each time you regenerate a response, the outcome will vary. These outputs are rooted in the concept of randomness, contributing to the uniqueness – and ‘human’ quality – of AI-generated content. Therefore, it’s crucial to embrace this variability as a feature rather than a flaw. Minor alterations in prompts can lead to diverse outcomes. For instance, consider the distinct results from the following prompts:
In a sentence, why should companies implement digital asset management?
Companies should adopt digital asset management to efficiently organize, store, retrieve, and distribute their digital assets, enhancing collaboration and streamlining workflows across the organization.ChatGPT
In a sentence, why should institutions embrace digital asset management?
Institutions should utilize digital asset management to effectively organize, preserve, and share their digital resources, optimizing access and utilization for educational, research, or archival purposes.ChatGPT
Verification is Key. One significant challenge with generative models is their potential to generate information that is either inaccurate or entirely fictitious. Consider the example where ChatGPT assumes a sole survivor of the Titanic disaster, despite the actual survival of over 700 individuals.
Who was the sole survivor of the Titanic?
The sole survivor of the Titanic was a passenger named Violet Constance Jessop. […]ChatGPT
Large language models are far from static; they undergo continuous updates. Consequently, reproducing examples from earlier versions may become impossible. While many are familiar with ChatGPT, they might be surprised by the numerous iterations: 3.5, 4 (representing the most current version as of the time of writing). Awareness of the training date of the LLM in use is crucial, as results may reflect outdated information. Consider this instance from ChatGPT:
Will the 2024 Chevrolet Corvette E-Ray feature an electric motor?
I’m sorry, but I lack information about events or developments occurring after September 2021.[…]ChatGPT
Due to their training on human-written data, a significant portion of which reflects historical prejudices, LLMs are inherently biased. While being conscious of AI bias constitutes an important step in recognizing and tackling potential injustices propagated by automated systems, it is imperative that we go further. We must insist that the companies offering access to these models prioritize transparency, accountability, and ethical practices in their development.
AI’s Imitative Nature
Contrary to its name, Artificial Intelligence lacks genuine cognitive understanding and common-sense reasoning. Notably, it lacks contextual awareness. As previously mentioned, LLMs are trained to emulate patterns present in their training data. The abundance of data contributes to astonishingly human-like results. However, it’s important not to be deceived by this resemblance and to verify outcomes.
Rectify this messy metadata:
Subject: Joe Blow
Date: Augu, 15, 203
Subject: Joe Blow
Date: August 15, 203
Manage Your Expectations and Put in the Work
While LLMs effortlessly generate textual content, they make assumptions based on their training. When querying ChatGPT, even though it may furnish details, these might not align with your expectations, or worse, could be fabricated. LLMs are akin to autocomplete engines. They deconstruct your prompt and, drawing from their training, complete the response. This process appears almost magical, yet it’s rooted in science and probability. Regenerating the response yields varying outcomes, which can pose challenges when aiming to fine-tune or specify results. You reap what you sow. Consistently achieving desired output demands time and patience. Precision in crafting prompts correlates with result quality. Successful interactions hinge on formulating prompts that yield desired responses.
Stay informed about LLMs, and temper your expectations regarding their capabilities. Generative AI, when complemented by checks and balances like manual quality assurance and automated validation, can prove a potent tool. In the metadata cleaning example above, offer the LLMs explicit instances to align with your expectations. Ultimately, bots excel with precise instructions but still require oversight.
Large language models wield significant potential but demand comprehension of their intricacies and limitations. Embrace AI’s evolving nature and employ it judiciously. Responsible usage, output verification, and consideration of AI bias are paramount.
David Shapiro ~ AI (Director). (2022, July 8). Prompt Engineering 101: Autocomplete, Zero-shot, One-shot, and Few-shot prompting. https://www.youtube.com/watch?v=v2gD8BHOaX4
‘The Ezra Klein Show’. (2023, January 6). Opinion | A Skeptical Take on the A.I. Revolution. The New York Times. https://www.nytimes.com/2023/01/06/opinion/ezra-klein-podcast-gary-marcus.html