 |
A model is used to represent or perform something. In AI, intelligence is represented or performed by an AI model.
AI model = algorithm (structure) + training parameters / learning values (tokens)
To train the model, we need to feed it genuine data or real data. If we train it with fake or poor-quality data, it will not perform well and cannot produce accurate results for the user.
Sometimes there are several models (multi-models). Different models use different algorithms and different tokens. GPT models are generative pre-trained models, meaning they are trained in advance and then generate outputs from their learned values. After training, they do not continue learning in real time—they simply perform using what they have already learned.
An AI model performs intelligence according to its learning values. These learning values can be thought of as its construction of realization or its level of experience. The algorithms are logical calculations.
In other words:
AI model = algorithms (logical calculations) + learning values / tokens (experience level)
Algorithms (logical calculations) are used for learning, thinking, and adapting data.
Examples of Algorithms:
- Linear Regression
- Decision Trees
- Neural Networks
- Support Vector Machines (SVM)
- Genetic Algorithms
- K-Means Clustering
Uses of Algorithms:
- Learning from data: Many algorithms are data-driven.
- Pattern recognition: Can detect complex patterns in images, video, sounds, language, or even mouse/touch clicks.
- Decision making: Can select the best path from multiple possible paths.
- Prediction: Can analyze and calculate the next steps.
The learning outcomes (or tokens/parameters) represent the experience level of an AI model. Some large-scale AI company models have more than 1 trillion parameters, such as ChatGPT and DeepSeek.
Algorithms in ChatGPT
There are 14 major algorithms/components in ChatGPT-like applications:
- Transformer architecture – the base design for GPT models.
- Self-Attention mechanism – lets the model focus on relevant words in context.
- Multi-Head Attention – multiple attention “views” combined for richer understanding.
- Feed-Forward Neural Networks – applied after attention in each layer.
- Positional Encoding / Embeddings – represent token order since attention itself is orderless.
- Layer Normalization – stabilizes training and inference.
- Residual Connections – prevent vanishing gradients by shortcutting outputs.
- Byte-Pair Encoding (BPE) Tokenization – converts text into tokens for the model.
- Softmax function – turns raw outputs (logits) into probability distributions.
- Cross-Entropy Loss – main loss function used during training.
- Stochastic Gradient Descent (SGD) variants – optimizer algorithms (Adam/AdamW) used to adjust weights.
- Reinforcement Learning from Human Feedback (RLHF) – aligns responses with human preferences.
- Proximal Policy Optimization (PPO) – a specific reinforcement learning algorithm for fine-tuning.
- Sampling/Decoding strategies – how the model chooses words (greedy, beam search, top-k, nucleus/top-p sampling, temperature).
Multi-Model Applications in AI
AI applications often use multiple specialized models, divided by task domains such as:
- Text
- Video
- Audio
- Images
These models can also be customized for specific tasks based on end-user requirements. Such customized systems are often called AI agents.
Sometimes, models are upgraded by:
- Increasing the number of tokens/parameters (to improve experience level), and/or
- Adding new algorithms (to enhance reasoning or capabilities).
0 Comments