Generative AI is a type of artificial intelligence that can produce new content including text, images, audio, and synthetic data. Large Language Models (LLMs) and Generative AI intersect, and they are both a subset of deep learning. But what exactly is an LLM?
At a high level, LLMs refer to large general-purpose language models that can be pre-trained and then fine-tuned for specific purposes. Let's break down exactly what that means.
🧩 Deconstructing "LLM" :
• Large ➡️ This refers to a large training dataset. It also refers to a large number of parameters, which are often called hyperparameters in machine learning. Parameters are basically the memories and the knowledge that the machine learned from the model training. Because of the huge datasets and tremendous number of parameters required, only certain organizations have the capability to train these models.
• General Purpose ➡️ This means the models are sufficient to solve common language problems across industries. This approach works because of the commonality of human language regardless of specific tasks, and it helps overcome resource restrictions.
• Pre-trained & Fine-tuned ➡️ You pre-train a large language model for a general purpose with a large dataset. Then, you fine-tune it for specific aims with a much smaller dataset.
🏆 Why are LLMs a Game Changer?
*The benefits of using large language models are straightforward: *
• A single model can be used for different tasks, including language translation, sentence completion, text classification, and question answering.
• They obtain decent performance even with little domain training data, allowing them to be used in "few-shot" (minimal data) or "zero-shot" scenarios (recognizing things not explicitly taught before).
• The performance of large language models is continuously growing when you add more data and parameters.
⚙️ How They Work: The Transformer Workflow:
• LLMs are almost exclusively based on transformer models. A transformer model consists of two main parts:
• [ Input Sequence ] ➡️ [ Encoder ] {Encodes the input sequence} ➡️ [ Decoder ] {Learns how to decode the representations for a relevant task}
🍦 The 3 Flavors of LLMs:
*There are three main kinds of large language models: *
• Generic Language Models: These predict the next word based on the language in the training data. Think of this model type as an autocomplete in search. For example, if the input is "the cat sat on", the model determines that "the" is most likely the next word.
• Instruction-Tuned Models: This type of model is trained to predict a response to the instructions given in the input. Examples include asking the model to "summarize a text," "generate a poem," or "classify text into neutral, negative, or positive".
• Dialogue-Tuned Models: This model is trained to have a dialogue by the next response. They are a special case of instruction-tuned models where requests are typically framed as questions to a chatbot. They are expected to be in the context of a longer back-and-forth conversation.
📝 The Power of Prompting:
How you talk to an LLM dictates the quality of what you get out of it.
• Prompt Design is the process of creating a prompt tailored to the specific task. For example, if you want a system to translate English to French, the prompt should be in English and specify that the translation should be in French. Prompt design is a general concept and is always essential.
• Prompt Engineering is the process of creating a prompt designed to improve performance. This may involve providing examples of the desired output or using effective keywords. Prompt engineering is a more specialized concept, only necessary for systems requiring high accuracy or performance.
Pro-Tip: Utilize Chain of Thought reasoning. This is the observation that models are better at getting the right answer when they first output text that explains the reason for the answer.
🛠️ Customizing Your AI: Tuning Workflows
A model that can do everything has practical limitations, but task-specific tuning can make LLMs more reliable.
[ Base Foundation Model ] ➡️ [ Domain Adaptation ] {e.g., Vertex AI tuning models specifically for legal or medical domains}.
If you need deeper customization, you have two distinct paths:
-
Fine-Tuning: Bring your own dataset and retrain the model by tuning every weight in the LLM.
The Catch: This requires a big training job and hosting your own fine-tuned model, which is expensive and often unrealistic.
-
Parameter-Efficient Tuning Methods (PETM): Tune a large language model on your own custom data without duplicating the model.
The Benefit: The base model itself is not altered. Instead, a small number of add-on layers are tuned, which can be swapped in and out at inference time.
☁️ Building with Google Cloud:
*If you want to move from theory to building, Google Cloud provides several powerful tools: *
• Vertex AI Studio: Helps developers quickly explore, customize, and deploy generative AI models. It provides a library of pre-trained models, fine-tuning tools, and a community forum.
• Vertex AI Agent Builder: Build chatbots, custom search engines, and digital assistants with little or no coding and no prior machine learning experience.
• Gemini: A multimodal AI model that is incredibly adaptable and scalable. Unlike traditional language models, Gemini is not
limited to text; it can analyze images, understand audio nuances, and interpret programming code.
• Model Garden: A resource that is continuously updated to include new models.
Top comments (0)