Large Language Models

What are they

Large Language Models (or LLMs for short) are Machine Learning Models that:

• are really good at understanding human language
• are made out of a huge number of parameters 

Why they exist

The main way people exchange information is through language. The most common way to record information is through writing. And ain’t nobody got time to go through all them blog posts and books and flyers and cave writings in one lifetime. 

In come the LLMs and their ability to store complex relationships between words. With those LLMs we can then: 

• automate tasks
• go through unstructured data
• or have just a chat with

How they work

Broadly speaking, LLMs are optimized to discover and remember statistical patterns given the huge amounts of text they are exposed to during the training phase. The more often certain words appear together, the stronger the connection the model forms between them. And the more parameters a model has, the more complex relationships it can remember.

Disclaimer

• This is an over simplification of what and how LLMs learn. 
• The training process itself is different based on the type of LLM.
• Please consult your local ML Engineer before training your own LLM.

Types of LLMs

Language Models come in different shapes and sizes depending on their purpose. To name a few, you have:

• Generative Language Models - generate new text based on a given input prompt by predicting the next word in a sequence
  • Example: GPT (Generative Pre-trained Transformer), Llama, Claude
  • Use Cases: Chatbots, AI agents, code generation, content creation, question answering

These are typically unidirectional, meaning they generate tokens one at a time in a left-to-right fashion.

• Masked Language Models - predict missing or masked tokens within a sentence
  • Example: BERT (Bidirectional Encoder Representations from Transformers)
  • Use Cases: Text classification, named entity recognition (NER), sentence similarity, embeddings

These are bidirectional, leveraging both left and right context, which makes them great for understanding tasks rather than generation.

• Sequence-to-Sequence (Seq2Seq) / Encoder-Decoder Models - encode an input sequence into a latent representation and then decode it into an output sequence.
  • Examples: T5 (Google), BART (Facebook), MarianMT
  • Use Cases: Translation, summarization, text-to-text tasks

These are flexible for tasks where the input and output are both text, often of different structure or language.

Sure, you can use ChatGPT (generative models) for most tasks. But in case you want a more cost-efficient model, that you can train yourself on your custom data and avoid dealing with hallucinations, you may want to keep your mind open for other options.

What Next?

LLMs are freaking everywhere these days and - as we say in Romania - people are gobbling them like hot bread. That’s why I think it’s important to not only know how to use them but also how they work.

Generative Models, for instance, are very resource-intensive and have the habit of surprising you with your pants down when you over-rely on them. 

But to get to the point, I feel everyone should know:

• how to prompt Generative Models for their full potential 

  • by learning the basics of prompt engineering - the better the prompts, the time and money you waste using ChatGPT

• what’s prompt injection and how to protect yourself from it

  • especially if you want to build your own user-facing product that uses LLMs
  • you can check out my game Prompt-it! if you wanna get a taste of that

• how to make Generative Models produce more reliable outputs 

  • be it through parameter tuning
  • developing tools that the models should use
  • or giving them access to a RAG-based system so they can make cheat-sheets

If you’re curious to know more or to take a glance of what things you can build using LLMs, here’s mindmap I built that may help.