𝗟𝗟𝗠𝘀 𝘃𝘀 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗔𝗜
Generative AI is a broad term. LLMs are a specific type of model within that group. You need to understand the technical differences to build better systems.
LLMs use Transformers to predict the next token. They work through autoregressive processes.
How an LLM works:
- The model takes a prompt.
- It converts text into tokens.
- It calculates the probability of the next token.
- It repeats this loop to form sentences.
Diffusion models work differently. They create images by reversing noise.
How a Diffusion model works:
- The system starts with random noise.
- It removes noise step by step.
- It follows a pattern to reveal an image.
- It uses prompts to guide the denoising.
A quick comparison: • LLM: Uses autoregressive Transformers. Output is text tokens. • Diffusion: Uses iterative denoising. Output is pixels. • GAN: Uses adversarial training. Output is synthetic data.
The boundary between these models is changing. Most modern systems now use Transformer backbones. Architecture is becoming compositional. This means different model types will work together in one system.