Why Using Language Models Is the Next Evolution in Human Expression
Communicating is Hard
There was a time when humans couldn’t rely on words — when a flicker of the eye or a simple sound was the only way to signal hunger, danger, or affection. Early humans communicated through gestures, expressions, and even the rhythm of footsteps in the dirt.
As our brains evolved, so did our desire to convey increasingly complex thoughts. Hand signals became symbols, and symbols transformed into spoken language. This leap allowed us to communicate abstract emotions, stories, and ideas. Language became the bedrock of human society, enabling cooperation on a scale far beyond any other species. In essence, language emerged as a social innovation — a bonding, teaching, and collaboration tool — through which we built societies, preserved knowledge, and created art.
Yet, for all its power, language remains a limiting factor in human expression. There are times when words fail us; when the complexity of our experiences or emotions seems to exceed the language we have at hand. Even with centuries of refinement, language struggles to capture the full range of human experience. This limitation is further aggravated by gaps in education and cultural differences, creating barriers in how we communicate and understand one another.
But now, innovations in large language models (LLMs) have opened the door to new possibilities in how we communicate. These models are rapidly integrating into our everyday lives, reshaping our ability to articulate complex thoughts and emotions. While they may not expand language’s inherent capacity to capture every nuance of human experience, they offer a powerful means to refine how we express ourselves. With thoughtful use, language models can help us find words that align more closely with our feelings and structure ideas for genuine understanding.
Demystifying Language Models
Before exploring how we supercharge our ability to communicate through LLMs, let us first take some time to understand how they work from a conceptual level. To do so, we’ll begin with a high-level look at machine learning. Avoiding technical details and nuance, we can think of machine learning as an implicit learning process: a computer is given a series of labeled inputs, which it uses to recognize patterns and develop general rules, or heuristics, about the data. After forming these heuristics, the model’s accuracy is tested on a separate set of inputs it hasn’t encountered before. Based on how well it predicts outcomes of unseen data, it adjusts its heuristics, iteratively refining them over multiple training cycles.
This process is not so different from the way humans learn. Just as we learn by encountering examples, making mistakes, and adjusting our understanding based on feedback, machine learning models “learn” by repeatedly testing and refining their interpretations of data until they reach a human-satisfactory level of accuracy.
These heuristics, or learned patterns, are captured in a set of numbers called weights. During training, the model adjusts these weights to represent the strength of various relationships within the data. Once the model is deployed in real-world applications, however, it doesn’t continue learning automatically. Instead, a complex mathematical operation is performed against the stored weights to produce predictions or responses based on new inputs. In essence, these weights serve as a static snapshot of what the model has learned, allowing it to apply its trained knowledge efficiently without further adjustment.
It’s a bit like someone who’s gone through years of schooling but is now forever stuck with what they’ve learned — no more new lessons, just applying the knowledge they’ve got, for better or worse.
With this foundation in mind, let’s take a closer look at language models specifically. Like other machine learning models, LLMs begin by processing massive amounts of text data to learn patterns, structures, and relationships within language. During training, words and phrases are transformed into numbers through a process called tokenization, allowing the model to interpret language mathematically. Each token is represented as a vector, or a list of numbers, capturing aspects of the token’s meaning. Think of each vector as a point in a multi-dimensional space — sort of like placing words on an extremely complex, invisible map where the directions aren’t north, south, east, and west but rather aspects of meaning like context, similarity, and usage.
This is called the embedding space. The embedding space allows the model to represent relationships between words in a mathematically meaningful way. For instance, words with similar meanings, like ‘dog’ and ‘puppy,’ will appear close together in this space, while unrelated words like ‘cat’ and ‘bank’ will be farther apart. What’s fascinating is that embeddings can also be combined or manipulated to reveal complex relationships between words. For example, if you take the vector for ‘king,’ subtract the vector for ‘man,’ and add the vector for ‘woman,’ the result is a vector very close to ‘queen.’
This shows how the model captures the underlying patterns and associations between concepts in a numerical form, allowing it to understand and navigate meaning in a nuanced way.
But capturing meaning isn’t as simple as positioning words on a flat map. The embedding space is n-dimensional, reflecting the endless complexities of language and meaning. It can be difficult to imagine this kind of space because, as humans, we’re used to a 3-dimensional world of length, width, and height. In an n-dimensional space, however, there are countless axes along which words can relate to one another, each representing a unique aspect of meaning.
To help visualize, imagine a word like ‘apple.’ In a 3D space, you could only place it based on three attributes, perhaps its physical characteristics like color, shape, and size. But in an n-dimensional space, you could position ‘apple’ based on infinite layers of meaning: its association with food, technology, symbolism, history, cultural references, and so on. Each dimension is a new axis, capturing these unique associations.
This n-dimensional nature is what allows the model to position words in a way that reflects their multi-faceted relationships. Words that are close to each other in this space share similar meanings in a particular context, but the model can also access other dimensions of meaning depending on the input. In this way, embeddings capture and navigate the incredible depth of human language and allow models to respond in nuanced ways.
Now that we have a sense of how language models represent meaning, let’s look at what these models are trained on and what they’re trained to do. During training, models are typically exposed to vast, diverse datasets that include books, articles, websites, and other text sources. This gives the model a broad understanding of language, topics, and the types of relationships between words and concepts. But what exactly is it learning? At its core, the model is trained to perform next-token prediction: given a sequence of words, it learns to predict the most likely next word, refining this skill over countless examples. This approach allows the model to construct coherent and contextually appropriate sentences based on statistical patterns in the language it’s learned.
Beyond this foundational training, language models undergo a second layer of refinement called Reinforcement Learning from Human Feedback (RLHF). Unlike the traditional next-token prediction, RLHF involves humans who evaluate and guide the model’s responses, adjusting them based on how well they align with desired outcomes. This step is crucial for what’s known as AI alignment — ensuring that the model’s outputs align with human values, expectations, and norms.
However, RLHF is also a double-edged sword, as it’s the stage where human bias is most likely to be introduced. By providing direct feedback, humans are, in effect, placing their ‘thumb on the scale,’ nudging the model to prefer certain types of responses over others. While this helps align the model’s output with user expectations, it also means that biases from human trainers can subtly influence the model, affecting the neutrality and balance of its responses. This human touch in training, though necessary for alignment, remains one of the key challenges in creating fair, balanced language models.
Once the model’s weights are fully trained and adjusted through RLHF, it’s ready for deployment and can begin interacting with users through a user interface. When a user inputs a query, their text is first tokenized, turning each word or phrase into a series of numbers that the model can interpret within its n-dimensional embedding space. This transformation enables the model to process complex language, map the input to relevant areas in its learned space of meaning, and generate an appropriate response.
To guide this response and set the model’s tone, system prompts play a crucial role. These prompts are developer-defined instructions that establish the model’s ‘personality’ or purpose, shaping how it should behave in each interaction. System prompts emerged to give models greater consistency and alignment with user needs, making them adaptable to specific applications. For example, a customer service bot might have a system prompt like, ‘You are a helpful and friendly assistant who answers customer questions clearly and politely.’
Importantly, these system prompts are invisible to the user, operating in the background to shape responses without explicit cues. When using a language model, it’s helpful to know that almost all consumer-facing applications — like ChatGPT — rely on these guiding prompts. They play a foundational role in determining the style, tone, and type of information the model provides, creating a more consistent experience for users while remaining hidden behind the scenes.
Altogether, language models are built from a layered architecture, where each component — from foundational training data to the final user query — contributes to the model’s ability to generate relevant, nuanced responses. Starting with vast knowledge bases, the model builds a mathematical framework in an n-dimensional embedding space to capture intricate relationships between words. It refines its weights through iterative learning and aligns responses with human expectations through feedback, while system prompts set the tone and purpose of interactions to make each experience feel uniquely tailored. This hierarchy, as shown above, enables language models to transform raw data into meaningful language.
Leveraging Language Models
Some approach language models like ChatGPT as a quick fix, thinking, “Just write the essay for me,” and simply feeding it a prompt. At the same time, others are so concerned about AI’s influence that they see any use of it as a threat to authentic expression, fearing that humans will stop thinking for themselves. However, both perspectives overlook the real potential of language models. Used thoughtfully, these tools can help us articulate ideas more effectively, serving as collaborators rather than shortcuts or intellectual crutches.
Because language models are trained on vast, carefully curated collections of text — including highly-regarded works and impactful journalism — they learn how language conveys complex ideas in ways that resonate across audiences. These texts, often the very ones people use to learn language themselves, allow models to become remarkably adept at structuring thoughts clearly and concisely. In fact, by tapping into methods the models have internalized for analyzing, contextualizing, and packaging language, we can often express ourselves with greater clarity and precision.
This understanding empowers us to use language models not simply as tools for churning out content but as collaborators in refining our ideas. When approached constructively, LLMs can help us find the right words, align our tone, and ensure our message lands well with others — all while staying true to our authentic voice. And with the right prompting and interaction techniques, we can shape our ideas, spark creativity, and explore fresh ways of communicating. Let’s explore some of these techniques to better harness the unique strengths of LLMs.
Context Loading
Before you even start prompting, try loading the context window — the space where the model ‘reads’ prior input — with information relevant to your goal. The context window functions as the model’s short-term memory, containing the initial information and setup that the model uses to frame its responses. By pre-loading this space with related ideas, background details, or even a stream of thoughts, you can prime the model to focus on the themes or directions that matter to you.
What’s powerful about context loading is that it doesn’t have to be structured. Since the model’s embedding space is governed by meaning-based relationships rather than strict grammar or diction, even a loosely organized stream of thoughts can effectively steer the model. For instance, you might jot down a list of related concepts, incomplete sentences, or even rough ideas — this raw input can serve as an effective foundation for the model to build on, orienting it toward specific regions in the embedding space that align with your topic or mood.
This approach not only makes it easier to convey complex thoughts but also reduces the time and effort needed to polish your input. By providing a “thought dump” directly into the context window, you let the model connect the dots within its embedding space, translating your ideas more seamlessly and at a lower cognitive cost.
Meta Prompting
Instead of simply stating your objective, meta prompting involves specifying how you want the model to reach that objective. This could mean asking for an analytical breakdown before forming a conclusion, adopting a contemplative tone rather than direct, or structuring a response in a step-by-step format. By instructing the model on the approach, you aren’t just influencing the words it uses; you’re adjusting how it organizes and connects information within the embedding space to better reflect your intent.
Think of it this way: language models are trained to predict the next token based on patterns in a vast dataset, which includes a wide range of formats, tones, and structural styles. When you engage in meta prompting, you guide the model toward regions of the embedding space that align with a particular structure and tone, modifying how the content is presented. In this way, you’re influencing both what the model says and how it prioritizes different dimensions of meaning and presentation.
Iterative Prompting
You’re not limited to accepting the first response an LLM provides. With iterative prompting, you can shape and refine the answer by describing how you’d like it adjusted — whether that’s to elaborate further, simplify the language, or take on a different perspective. This process can be a dynamic, step-by-step chain of prompts that guides the model gradually toward the response you’re seeking, or it can involve directly editing your initial prompt in one go, depending on the tools available in the application you’re using.
In many cases, iterative prompting involves a sequence, or chain, of prompts, where each re-prompt includes the original prompt and response. This continuation of context allows the model to build upon its previous output, keeping elements of the initial answer while incorporating your adjustments. For instance, you might start with a high-level summary and then ask the model to delve deeper into specific points. Each prompt in this chain subtly shifts the embedding space focus, helping to refine the output until it aligns closely with your intent.
Alternatively, some LLM applications let you directly modify your initial prompt, so the model starts fresh with your refined instructions. This can yield different results, as the model has no prior context from previous interactions. While editing the initial prompt may provide a clean slate and a new perspective, the chaining approach allows you to leverage the model’s memory of previous answers, maintaining continuity in tone and content. Each technique has its advantages, and knowing when to use one over the other can help you achieve more nuanced, intentional results.
Permutative Requests
For deeper exploration, ask the model to generate responses in multiple permutations — expressing the same idea in varying tones, perspectives, or styles. This technique is relatively simple but powerful, as it allows you to access a broader slice of the embedding space and explore different facets of meaning and presentation. By using straightforward prompts, you can introduce a bit of entropy — randomness — in how the model interprets your request, helping you see a wider range of structural variations, from formal to conversational or analytical to imaginative.
While permutative requests are often used for exploring structural differences, you can also apply this technique to surface meaning-based variations. For example, you might ask the model to explain a concept like “freedom” in different contexts — political, philosophical, and personal — to see how each perspective alters the interpretation. This could look like:
- “Explain the concept of freedom as it relates to political governance.”
- “Describe freedom from a philosophical viewpoint.”
- “What does freedom mean on a personal level?”
By requesting these variations, you get a fuller view of how the model can articulate your concept, allowing you to choose the version that best suits your needs. Whether for style, structure, or nuance, permutative requests let you explore the model’s full capacity to express an idea across diverse dimensions, broadening your toolkit for effective communication.”
Bringing It All Together
At the heart of each technique we’ve explored lies a common thread: the art of navigating and manipulating the embedding space. Unlike human communication, where language flows intuitively to connect meaning, communicating with language models involves a unique layer of strategic prompting to access specific regions of this complex, multidimensional space. With LLMs, language becomes not just a message but a tool for directing attention, precision, and style within the model’s learned map of words and ideas.
Each technique — from meta prompting to context loading — allows us to adjust how the model interprets and assembles responses. While these strategies are effective individually, they become even more powerful when used in combination. For instance, you might load context to set a foundational tone, then refine with iterative prompting, or apply meta prompting and permutative requests to explore both nuanced meaning and varied presentation. Together, these methods transform LLM interaction from a simple Q&A exchange into a dynamic process of sculpting responses that are deeply aligned with your intent.
It’s in this way that language models have the potential to revolutionize human communication. On one hand, they drastically reduce the cost of articulating complex thoughts; on the other, they are inherently skilled at packaging language in ways that are broadly accessible. By applying prompting techniques like those we’ve covered, you can elevate your communication far beyond those who simply offload tasks to AI. This approach allows you to inject your human touch into the final product while letting AI handle the heavy lifting of crafting language.
However, unlocking this potential requires an understanding of how the model truly works beneath the surface — how embedding spaces, training data, and prompting techniques interact to produce coherent responses. But with this knowledge, the results can be remarkably powerful. For example, this entire article was created using an LLM. I brought the ideas, prompting techniques, and my understanding of machine learning and language models, but ChatGPT helped me flesh out a base outline, expand sections, brainstorm different title permutations, and put the meaning I had in mind into well-structured sentences. Much of the content came from stream-of-thought entries that I loaded into the context window, letting the model refine and structure my ideas for a clearer presentation.
In this way, I could focus on my message and meaning without wrestling over each word. Hopefully, this approach helped package the article in a way that’s easy to understand and engaging for you, the reader.
For a behind-the-scenes look, you can view the origin of this article in my conversation with ChatGPT.
