What are AI Models?

AI models is a program or algorithm that has been trained on large data to recognize the patterns, make better predictions and perform specific tasks from the input data without the intervention of human.

Example: Think of all the models you see these days from ChatGPT to Spam Checker, all are the AI models at different stages ranging from heavy(Advanced) to basic models. We’ll cover how they are evolved and trained in short with examples.

How does these models actually learn?

You’ve got the right question, As we just knew that they are trained on large data, Let’s see how they had been trained. This training involves the same technique similar to us as a human from child to adults learn. We learn the things in three different ways through guidance, self and experience/suggesstions.

• Supervised Learning - You give(train) the model input with answers (called labels), and it learns to map input → output.

  Example: Training a model to detect spam by giving emails labeled “spam” or “not spam”

• Unsupervised Learning - You give(train) the model only input data without answers, and it finds hidden patterns or groupings on its own.

  Example: Grouping Retail customers into segments based on their spending behavior — without telling the model about their spending nature or power.

• ReInforcement Learning - You give the model input data and the model learns by trial and error -getting rewards for good actions and penalties for bad ones. Often paired after supervised learning to improve behavior.

  Example: A robot learning to walk: tries moving forward, falls, gets negative reward; eventually figures out balance by optimizing actions.

  

What makes these models so powerful?

A special kind of architecture called “Transformer” is behind the success of the many AI tools like ChatGPT. A Transformer is a model design that helps AI understand context in data, especially in language. Instead of reading data one word at a time like older models, Transformers read everything together and decide what to focus on using a method called attention.

Example: “The Elephant was so huge that it doesn’t fit in the cave”, a Transformer helps the AI decide whether “it” refers to the elephant — by paying attention to context.

From where these LLMs came from?

Once these transformers evolved, the real game started in the area of AI, where the transformers are scaled up with huge data and parameters to create LLMs. Parameters are internal values (like weights) the model learns during training. Parameters determine how the model behaves on new data. A Large Language Model(LLM) is a model trained to understand and generate human language. These LLMs are trained using supervised and reinforcement learning on huge amounts of text from books, websites, and code.

Example: ChatGPT, Bard, Claude as they understand the human language and respond to it accordingly.

How Can LLMs Perform So Many Tasks?

Here comes Shot Learning the reason behind LLMs performance capabilities.

Zero-shot, One-shot, Few-shot Learning

Unlike traditional models, LLMs don’t need to be retrained every time you give them a new task. They can often perform tasks with:

• Zero-shot learning: You ask a question without showing any example.

  Example: "Translate ‘Bonjour’ to English." It works (zero-shot) and gives “Good Day”

• One-shot/Few-shot: You give one/few examples.

  Example: "Translate the following: Hola = Hello. Ciao = Hello. Namaste = ?" That’s few-shot.

How to improve or use these LLMs?

Eventhough LLMs are trained generally, you may need few places where the user might ask for specific data or answer in few industries like Healthcare, Policies followed. For these use cases, We can fine-tune LLMs on specific tasks/use-cases by giving them extra training on focused data. Think of a situation where, A company may fine-tune GPT on its internal customer service conversations so it understands company-specific tone and vocabulary.

Sometimes, We don’t require the answers generated by the model to restrict to one industry or a topic. Instead of retraining or fine-tuning the model, sometimes all you need is to ask it the right way and get the appropriate or relevant answer from the model. This is where Prompts play a crucial role. Prompt Engineering is the art of crafting the right input to get the best possible output from an AI model.

Example: Instead of asking the model just - “Explain me Inflation”, Ask something - “Explain Inflation like I’m a 10-year-old who loves toys.” This gives more context and guides the tone.

Final Thoughts:

Everything in AI — from chatbots to image generators — starts with a model. Understanding Supervised, Unsupervised, ReInforcement Learning defines how a model is trained initially. Most LLMs like GPT start with supervised learning. Understanding ReInforcement Learning with Human Feedback gives you idea of how LLMs are less likely prone to misinformation.

Transformers are the backbone of today’s most advanced models. As LLMs contain billions of parameters, the model learns and are trained on massive datasets. The performance and flexibility comes from the training scale using shot learning and the Transformer architecture. Replaces traditional fine-tuning in many user-facing tasks depends on the model parameters and pre-training.

Prompt engineering is especially useful in few-shot learning, where you provide examples in your input prompt itself and expecting the model to provide the answer/solution based on your requirements or context instead of giving a general answer.