What Are Large Language Models?

Key Takeaways

One of the biggest topics around AI is how to model intelligence artificially. While the word intelligence can be an elusive concept, this is a critical question to answer, especially if you want to create AI models that can demonstrate intelligence in meaningful ways. For example, a robot can observe and operate in the physical world, is able to demonstrate spatial intelligence and temporal intelligence by picking up objects, describing the physical world, and carrying out meaningful tasks that are typically done by humans. Another form of intelligence is in language. And unlike physical world that robots need to face, languages are higher abstraction that describe the physical world. Which means demonstrating intelligence in human language is a different kind of intelligence than spatial intelligence or temporal intelligence that has more to do with time and space. Currently, we use large language models to approximate the linguistic intelligence by training them with massive amounts of text to learn patterns in languages like grammarss, syntax, and semantics. So just like how robots need to understand the laws of physics, LMS need to understand the laws of linguistics, large language models today can be anywhere from few billions to trillions of parameters in size, which means we need to have dedicated graphic cards, which means some models require hundreds of thousands of investments to run a trillion parameter model for one person. Popular models like GPD 5.2, Opus 4.6, 6 and Gemini 3.1 are all under the classification of a large language models that we interact with. Around 2023, we started to see different flavors of LLMs where now they can also accept other modalities like images, videos, and audios which gave birth to a type of LLM called multimodality. And the line between traditional LLMs that only understand languages has extended to understanding space and time by being trained to take tokens in other modalities as well. Around 2024, OpenAI also released what's called reasoning model where the LLM demonstrates thinking by essentially searching for the best answer instead of only generating the first likely response in a single pass. As you can see, the proliferation of different types of LLMs is still beginning where variants like omni models where it can simultaneously stream audio and interact in multiple modalities all at once just goes to show how far we can extend a simple LM to carry out tasks that go beyond simple chat interactions to a more useful tasks like real-time translation between different languages, voice assistance, and coding with screenshots and terminal logs. So, while we're still early in our ability to capture intelligence artificially, the real question that we should really ask ourselves