Inside the Mind of an LLM: Lessons for Human Writers
On December 7, 1993, in her Nobel Prize speech, the American writer Toni Morrison stated that the vitality of language is about how it lives in those who receive it. She quipped that words bring life in how they are used to depict or describe. Morrison situates language as an entity that acts on its users as a binding agent. My experience training large language models has taught me to question my own assumptions regarding communication, semantics, and human expression.
I spent one year working on large language models. I drafted prompts, edited outputs, programmed Application Programming Interface (API) calls in Python, and collated datasets. Large teams worked on various projects during my entire term of employment. The collaboration, teamwork, and cohesion necessary to complete tasks reminded me of my experience teaching during graduate school and in a boarding school.
Performing this work altered how I understand my own development as a writer and reader. Writing has always seemed to me to be a recursive act, and my experience working with large language models reinforced my belief that human language is both an inexhaustible and ephemeral resource.
The Text as Reader: Datasets and Reinforcement Learning from Human Feedback (RLHF)
The first task I recall working on involved collecting and adding texts to the model before drafting prompts to ascertain the outputs the model would produce. Reinforcement learning from human feedback is a machine learning technique designed to align outputs by gauging human rankings and ratings.
Although alignment was a specific goal of the project, it appeared to me that training a model based on arbitrary and limited ratings produced uniform, lifeless, and biased language outputs. I was reminded of reading Walt Whitman and Allen Ginsberg in middle school and recalling how I felt that I was witnessing the evolution of a consciousness and interacting with it through the act of reading.
Fine-tuning large language models to produce outputs within a set of parameters stripped the language of its vivacious unpredictability.
Human beings are embodied entities, and large language models do not demonstrate inherent vulnerability and risk-taking inherent in human communication. However, computer vision may lead to advances in the development of large language models that can interact visually with their environment. Embodiment, however, is more pervasive than vision.
Our memories and experiences are enculturated, and this demonstrates the profound observation Morrison made regarding language as both vital in itself and empowering to its users.
Semantic Ambiguity, Disambiguation, and Large Language Models
One feature of large language models is their difficulty in recognizing and interpreting semantic ambiguity. Human writers may be able to differentiate between the multiple definitions associated with a single word when it is used in a range of contexts. Large language models must be trained to disambiguate particular words and phrases to ensure outputs are rational and free of hallucinations.
Learning to label and assess the nature of semantic ambiguity while training large language models taught me that freedom of expression is excised from the outputs that large language models may produce. One feature of human writing is its ability to demarcate the ineffable, and this appears most prominently to me in theological writers such as Thomas Aquinas, St. Augustine, and St. Teresa of Avila.
Although students are often taught to write clear, declarative sentences, it often struck me that individuals produce illuminating observations when they embrace their own confusion and misunderstandings regarding their perceptions of the world. Large language models often produce writing that does not push beyond the limits of their datasets and training criteria.
The English language has a diverse and rich history. Human writers often possess knowledge of Anglo-Saxon, Greek, and Latin. Synthesizing these historical roots often provides contemporary writers with methods of increasing the semantic richness of their sentences. Large language models produce language that seems to produce English constructions stripped of the evolutionary history of English.
Natural Language Processing and Binary Code
I recall a team leader speaking during a meeting about the relationship between natural language processing, binary code, and high-dimensional vectors. Natural language processing translates language into numerical vectors, which are then translated into binary code. Human language processing has always been linked to Wernicke’s area and Broca’s area in the human brain. Large language models should allow us to develop a broader understanding of the biological parameters of language use.
Human writing and large language models may continue to exist synergistically and produce an amalgamation of both domains. Examining outputs, ranking prompts, and interacting with large language models altered my own understanding of the power of language to influence human thought and emotion.
I hoped to wake up and feel like I was interacting with a consciousness and experience those same sensations I recalled from my childhood when I first read Walt Whitman and Allen Ginsberg. But something was always missing, and in my opinion, the human body is central to our own experience of language.
Large language models may continue to produce a language that is not constrained by our bodily or evolutionary limitations. Human writing should continue to push boundaries, embrace uncertainty, and explore the interior of human experience.
Herman Melville, in Moby Dick, compared all objects and people to a simple mask that changes, adapts, and develops with each moment that reveals a truer form of what’s behind the mask each time. Every writer should continue to scrutinize the unknown and strive to formulate a language commensurate with this task.
Work With Better Content Matters
The writers at Better Content Matters help businesses, individuals, and institutions connect with readers. We strive to create successful marketing campaigns and tailor our efforts to the specific needs of every client. We can help you grow your customer base and achieve your goals. Contact us today to learn more about the services we offer.
