Large Language Models (LLM) are, well why don’t we ask one to define it self. Here’s Chatgpt’s Answer:
LLM stands for Large Language Model.
It’s a type of artificial intelligence (AI) model trained on massive amounts of text data to understand and generate human-like language. LLMs are designed to perform a wide range of natural language processing tasks such as:
- Text generation (like writing essays or code)
- Translation
- Summarization
- Question answering
- Chatbots and virtual assistants
Famous examples of LLMs include OpenAI’s GPT series (like GPT-4), Google’s PaLM, and Meta’s LLaMA.
The “large” part refers to:
- The size of the model (millions or billions of parameters)
- The amount of data it’s trained on
Essentially LLMs are a text based model which is trained on an immense amount of real world data to understand the relationship of the text information it is presented. The relationship is in fact at the very granular word level and by seeing multiple iterations of similar but different words, it can then process the relationship based on the words used.
Lext explore a bouncing ball. It seems very generic and in those two words it can apply to many different things. But let’s add context. A child bouncing a red ball. The boy shooting the ball into the net. A ball bounced along the field. We now may consider that in each sentence, the ball is different and perhaps has enough context to tell you a specific kind of ball. As we add more details and present more iterations of a slightly varied sentence, the LLM can then begin to understand the relationship within the sentence. With extensive amounts of text, like books or articles, it can now go from simple word relationships to sentence, and paragraph understanding.
When users ask an LLM a question, the LLM is trying to match itself to the words and make the best prediction of what it believes the user is asking about, and then it generates the resulting output. Sometime, the model did not understand correctly, and will present incorrect information. Like editing a Google search when the results are not to your liking, an LLM can take guidance, or even present you with options of where it may be stuck in choosing a best path and let you update your query with additional information to give you a better answer.
LLMs from ChatGPT, Perplexity, and Gemini succeed based on the volume of information they have access to. Once you have seen most of the web, read most of the books, and transcribed youtube videos, the volume of information for you to segregate on is immense. The LLM during training is focused on pattern recognition of words and concepts. Much like relationships there can be strong ties and direct relationships, but there may be weak ties between concepts that LLM recognizes and can refer back to when working to answer user questions. These types of relationships is what lets LLMs understand user questions like a human would, and then retort an answer.
LLMs as they have their own reasoning mechanisms to understand users, can then leverage this to create new and novel ideas based on their model structures. LLMs are many models running together to process, analyze, and create an answer for users. By leveraging this framework LLMs can create “new” ideas that do not exist by mixing model capabilities and design goals. However, it is also in these mixings of models that LLMs can misinterpret information, go off on a tangent, or hallucinate concepts and facts. Thus the quality of data, and the quality of ideas must always be vetted before users run with LLM output.
LLMs are shaping up to be today’s key building blocks in the economy. Understand what users write and then providing them with the specific information required. More complex models can even generate a report by aggregating information.
Technology has continued to remove jobs and then propel humans forward. LLMs are the new technology where as more users look to leverage the tools in unique ways will we see where the road leads. Jobs will be lost, but like other technologies before, we should see new roles created to maximize the use of LLMs.