In the vast ecosystem of artificial intelligence, one product that has captured both the imagination and skepticism of the public is OpenAI’s ChatGPT (and GitHub Labs). For those unacquainted, this model was designed with the prowess to simulate human-like text generation. Some critics have been quick to dismiss it as a mere “autocomplete” tool, or worse, a “bullshit generator.” However, a more astute understanding reveals that it functions more as a “first draft generator.” And in the complex world of artificial neural networks, this distinction matters.
The Human and AI Process of Sentence Formation
At the very heart of both human communication and AI-generated text lies the concept of “tokens.” Imagine a token as a fundamental unit of thought, typically words or phrases that hold semantic meaning. When humans communicate, they pick out a few of these tokens based on their thoughts, memories, and intentions. These selected tokens then act as a scaffold around which the rest of the sentence is built.
ChatGPT operates similarly, albeit driven by algorithms and vast amounts of data. It chooses tokens based on its training and the patterns it recognizes in the input. With these tokens as anchors, it predicts and generates subsequent tokens to form a coherent sentence.
The Publishing Analogy
I’ve come to these conclusions following using ChatGPT, and GitHub Labs, to generate larger works. My findings have been surprising, and I dispute these understatements of their capabilities.
To fully grasp the capabilities and potential of ChatGPT, one must first look at the intricate processes of the publishing world. An author’s manuscript undergoes several layers of refinement, from the rough draft to the final polished product ready for readers. The manuscript is edited for structure, flow, clarity, grammar, and numerous other parameters. It’s not just about producing text; it’s about refining and optimizing it.
Now, envision ChatGPT as a fledgling author presenting its first draft. Its work is impressive, far exceeding its human counterpart in both time and results. However, akin to any first draft, it lacks the refined touch that comes from rigorous editing. As with any first work, plot holes exist, and the entire work may feel dry or lacking emotion. Yet, given a few corrective edits, the work can start to shine.
The Shortcoming: Editing Larger Works
One of the significant limitations of ChatGPT, and most text-generating AI models, is their inability to manage, review, and refine larger bodies of work. Editing, in its true essence, requires the capacity to:
- Summarize portions of a text: An editor will often have to condense sections for brevity or clarity.
- Review and contrast different parts of the text: This involves ensuring consistency, flow, and thematic coherence.
- Maintain a memory of changes: Understanding why certain alterations were made and ensuring those changes reflect consistently throughout the work.
ChatGPT, as sophisticated as it is, doesn’t possess the integrated “memory” or the refined discernment of a human editor. It responds based on patterns in the data it was trained on, and while this leads to impressive text generation, the model lacks an overarching perspective on large pieces of content.
The Path to Superintelligence
Achieving superintelligence in AI entails not just producing text but comprehensively understanding, reviewing, and improving upon it. By integrating a mechanism where the AI can summarize its work, recall previous edits, understand the rationale behind changes, and use that knowledge to refine the overall product, we edge closer to the coveted realm of superintelligence.
But until that point, while ChatGPT might not replace the meticulous touch of human editors, it certainly emerges as a valuable tool in the creative arsenal, helping to scaffold ideas, produce initial drafts, and assist in the ideation process.
To dismiss OpenAI’s ChatGPT as mere “autocomplete” is to overlook its nuanced capabilities and potential. It’s a testament to where AI stands today and offers a tantalizing glimpse of where it could head tomorrow.