Large Language Models (LLMs)
Classification
(aka resistance to structural change)
NOTE: This classification applies to specific transformational depths (from seed boundaries). SOS Classifications cannot be compared across different depths.
So a “resilient structure” classification for astronomical bodies cannot be compared to one for human immunity series.
LLMs are brittle symbolic systems, held together by training coherence and system memory. They fail or mutate when inputs, hardware, or social framing shifts too far.
Type of boundary
Others
Understanding the boundary
Environmental context
A large language model is a tool. As such, and within the specific context of using it as a tool, LLMs operate at a Higher than human scales of reality. However, when viewed from a broader lens – LLMs have not yet reached the same level of complexity as life. This is why they are categorized as being ‘mostly Lower than human‘
Even a relatively uncomplicated animal like a wasp would operate at a higher scales of reality when you consider the complexity of all the functions that life solves for. Examples include, an immune system, a reproductive mechanism, the ability to take in multiple different types of input signals etc.
The specific environmental context during tool-use will consist of cloud computing infrastructures, data centers, and digital ecosystems. They interact with users, developers, and external datasets through API calls, user prompts, and machine learning pipelines. Their presence spans research labs, enterprises, and consumer-facing applications.
Mechanism for determining boundary
Like all tools – LLMs too have two complementary types of distinction mechanisms.
The first type of distinction is physical. Surprisingly for digital tools, the physical element is not as important as it is for other types of tools. But that doesn’t mean a physical boundary doesn’t exist – after all LLMs and any training data and computational methods will be stored on a server. They don’t exist in the ether.Â
The abstract (or biologically derived) component of the LLM boundary is much more important. It is defined by the parameters, architecture, and data it has been trained on. It is delineated by its training corpus, computational limitations, and the external inputs it can process. Unlike traditional software, its responses are probabilistic, meaning its boundary is functionally determined by statistical inference rather than deterministic rules.
Strangely enough, this also means that an LLM that is bad at predicting expected answers would not even qualify as an LLM – think of a random word generator but one that is trained with the best statistical methods to generate the most non-sequitur answers to a question.Â
Associated boundaries: higher scales
(not exhaustive)
- Artificial intelligence ecosystems: The broader AI research and deployment landscape, including computer vision, reinforcement learning, and other AI subfields.
- Human knowledge systems: The global body of human-written text, history, and culture, which informs the model’s outputs.
- Technological infrastructure: The computational and cloud-based networks supporting LLMs, including GPUs, TPUs, and distributed computing systems.
Associated boundaries: lower scales
(not exhaustive)
- Neural network architecture: The individual layers and nodes that process information within the model.
- Training datasets: The vast text corpora used to develop the model’s language capabilities.
- Tokenization and embeddings: The sub-word and word representations that allow the model to interpret and generate text.
Understanding adjacent boundaries (Biological types only)
Lower-fidelity copies
(not exhaustive)
NA
Higher-abstract wholes
(not exhaustive)
NA
Understanding interactions
Most commonly interacting boundaries
at similar scales (not exhaustive)
1. Users (Researchers, Developers, End-Users)
- Role: Send prompts, receive outputs, and provide feedback on responses.
- Timing: On-demand whenever someone types a query or integrates the model into an application.
- Symmetry: One-way—user gives text prompt; model returns text—but user feedback can guide future fine-tuning.
Â
2. Training Data (Large Text Corpora)
- Role: Provides patterns and examples that LLM learns from.
- Timing: Once during training; updated periodically when retraining or fine-tuning.
- Effect: Quality and bias of data directly affect model’s output accuracy and fairness.
Â
3. Compute Infrastructure (GPUs, TPUs, Servers)
- Role: Supply processing power and memory for both training and inference.
- Timing: Continuous usage during training (weeks to months) and on-demand for inference (milliseconds per query).
- Effect: Faster hardware speeds up training and reduces response latency; limited resources slow operations.
Â
4. External Tools and APIs (Databases, Knowledge Bases, Plugins)
- Role: Augment LLM outputs with up-to-date facts or specialized functions (e.g., math solver).
- Timing: Real-time calls during inference when model needs specific data or tools.
- Effect: Improves accuracy on niche tasks; adds complexity in coordinating prompts and tool results.
Mechanism for common interactions
(not exhaustive)
1. Prompt Processing (Tokenization and Encoding)
- How It Starts: User’s text prompt is broken into tokens (words or subwords).
- What Flows: Tokens convert into numerical embeddings that the model can manipulate.
- Effect: Sets the context for generation; poor tokenization can confuse the model or distort meaning.
Â
2. Attention Mechanism (Internal Information Flow)
- How It Starts: Model examines which parts of the prompt are most relevant at each generation step.
- What Flows: Attention scores guide how much weight to give to each token when predicting the next word.
- Effect: Allows the model to produce coherent and contextually relevant output, maintaining consistency across long prompts.
Â
3. Parameter Update (Fine-Tuning from User Feedback)
- How It Starts: Users label outputs as good or bad, creating a feedback dataset.
- What Flows: Backpropagation adjusts model weights during a fine-tuning step.
- Effect: Model improves on specific tasks or domains but may also overfit if feedback is narrow.
Â
4. Resource Management (Scaling Across Servers)
- How It Starts: Multiple GPUs or TPUs share the workload for large inference requests.
- What Flows: Data parallelism—each server processes a slice of data or a portion of the token sequence.
- Effect: Reduces response time for multi-user scenarios; balancing loads prevents bottlenecks.
Other interesting notes
- An LLM’s boundary is functionally shaped by it’s ability to predict what a user is looking for — trained not to understand, but instead use statistics and probability to provide the best likely answer for a given user. It lives on the edge of probability, not truth, and its fluency is a mirror without memory.
- Its paradox is that it speaks convincingly, but knows nothing — its power comes from pattern density, not perspective. Yet it reshapes how we write, think, and ask, even as it lacks any anchor in being. It mimics life without needing to be alive — reflecting our data back at us, sometimes in distorted form. It has no will, but it can persuade. No belief, yet it can generate conviction.
- LLMs remind us that boundaries don’t always contain content — sometimes they just route energy. And sometimes, the most convincing performance of knowledge is a function of scale, not insight.