Understanding LLMs

Understanding LLMs#

What Is a Language Model?#

A language model (LM) assigns probabilities to sequences of words. Models are trained — their probabilities estimated — using large collections of naturally occurring text (a corpus). A language model can be used to select the most probable next word given the preceding context.

What Makes a Model “Large”?#

A language model is considered large when:

  • It has billions (or approaching trillions) of parameters

  • It was trained on billions or trillions of words/tokens

As scale increases, new emergent capabilities appear — such as complex reasoning — that smaller models showed no indication of. This is why large language models have attracted so much research attention.

The Transformer Architecture#

Transformers are the key architectural innovation that enabled language models to scale. They allow massive parallelization of training and inference on GPUs. The foundational paper is Attention Is All You Need (Vaswani et al., 2017).

Pre-Training and Fine-Tuning#

Pre-trained models are trained via self-supervision on vast quantities of text. These are sometimes called foundation models. They can be used as-is or further trained (fine-tuned) to improve performance on specific domains or tasks.

The assistant-style models most people are familiar with have been further trained using supervised fine-tuning and reinforcement learning from human feedback (RLHF) to behave safely and helpfully — going beyond simple next-word prediction.

Generative Models#

Generative models produce content as output. “GPT base” models do simple next-word prediction. RLHF-trained assistant models respond to questions in a helpful, contextually aware way.

Choosing a Model: Open vs. Closed Source#

The choice between a closed-source API (like OpenAI or Anthropic) and an open-source model (like Llama or Mistral) depends on your research needs:

Consideration

Closed-Source (API)

Open-Source (e.g., Ollama on KLC)

Ease of use

Very easy — send a request, get a response

Requires setup on KLC

Performance

Frontier models (GPT-5, Claude) are state of the art

Slightly behind frontier, improving rapidly

Cost

Usage-based (per token); costs scale

Free to run on KLC compute

Data privacy

Data sent to external servers

Data stays on KLC

Reproducibility

Models update without notice

Pin exact version

Customization

Limited fine-tuning options

Full access to weights; fine-tune on your data

Use a closed-source API when:

  • You need state-of-the-art performance

  • Data privacy is not a constraint

  • You want the simplest possible setup

Use an open-source model when:

  • Your data is sensitive, proprietary, or IRB-governed

  • You need guaranteed reproducibility

  • You want to fine-tune on domain-specific data

  • Cost at scale is a concern

For the Kellogg-specific tradeoffs and how to run an open-source model on KLC, see Open Source LLMs.