Basic Concepts of Artificial Intelligence (AI)

AI systems learn patterns and rules by analyzing vast amounts of information (e.g., images, text, audio).

Once trained, AI applies its learned knowledge to make informed decisions or predictions about new data.

Data as the Foundation: AI systems need data—lots of it. This data can come from various sources, such as text, images, videos, or sensor readings.

Preprocessing: Before AI can learn, data must be cleaned and structured. Preprocessing involves:

• Removing noise (e.g., duplicates or irrelevant data).
• Normalizing data (e.g., scaling numeric values).
• Labeling data for supervised learning tasks.

Example: To train a model for image recognition, millions of labeled photos (e.g., "cat," "dog") are organized and formatted for input.

AI systems rely on mathematical models to interpret and analyze data.

The choice of model depends on the task:

• Linear regression models for simple predictions (e.g., sales forecasting).
• Deep neural networks for complex tasks like image recognition or language understanding.

Modern AI models like transformers (used in GPT and BERT) are particularly powerful for tasks involving sequences, such as language.

What Happens During Training: Training is the process where the model learns patterns in data by adjusting its parameters. This is done using machine learning algorithms like gradient descent.

Optimization Process:

• The model makes predictions based on the training data.
• An error function measures how far the predictions are from the actual answers.
• The model adjusts its internal settings (weights and biases) to minimize this error.

Computing Power Matters: Training can require massive computational resources, especially for large-scale models like GPT.

Once trained, the AI model moves to the inference phase, where it applies its learned knowledge to new data.

Example: A trained chatbot can generate responses to user questions it hasn’t seen before by inferring patterns learned during training.

AI systems improve through feedback. Users or engineers can provide corrections, and the model incorporates this new data into future training cycles.

Real-World Example: Recommender systems (like Netflix or Spotify) adjust suggestions based on user feedback, such as likes or skips.

Algorithms are step-by-step instructions that guide the AI system in processing data.

Popular AI algorithms include:

• Decision Trees: For simple decision-making tasks.
• Convolutional Neural Networks (CNNs): For image and video analysis.
• Recurrent Neural Networks (RNNs): For tasks involving sequences, such as speech recognition.

Inspired by the human brain, ANNs are layers of interconnected nodes (neurons).

Each neuron processes input, applies weights, and passes the result through an activation function.

Deep Neural Networks (DNNs) use multiple layers, enabling AI to solve complex problems.

AI learns in different ways based on the task and available data:

• Supervised Learning: The model learns from labeled data.
• Unsupervised Learning: The model identifies patterns in unlabeled data.
• Reinforcement Learning: The model learns through trial and error, receiving rewards for correct actions.
• Self-Supervised Learning: A hybrid approach where AI generates labels from raw data, often used in cutting-edge models like GPT.

Vectors: AI processes data as vectors—numerical representations that encode features of the input.

Embeddings: In tasks like language processing, embeddings are dense vector representations that capture semantic meaning (e.g., "dog" and "puppy" have similar embeddings).

Most modern AI systems start with pretrained models, which are trained on massive datasets.

Example: GPT models are trained on diverse internet text, giving them a broad understanding of language.

Fine-tuning adapts these pretrained models to specialized tasks using smaller, task-specific datasets.

Example: A GPT model can be fine-tuned to become a customer support chatbot for a specific company.

Knowledge learned in one domain is transferred to another, reducing the need for training from scratch.

Example: A vision model trained to recognize objects can be fine-tuned to detect specific items like medical anomalies in X-rays.

GPUs accelerate matrix computations, making them essential for training AI models.

• Rule of thumb: In example 8B Parameter LLM model (usually) requires GPU with 8Gb of VRAM.
• Bigger (and usually smarter) the model is, more VRAM you need in order to run the model.
• My personal recommendation is: NVIDIA GPU with 16Gb of VRAM (this would allow you to run most of the smaller models locally on your computer in decent speeds). Needless to say, more the merrier.
  

Specialized chips designed by Google for AI workloads, offering even faster processing.

Groq & the Language Processing Unit.
The Groq LPU, AI Inference Technology, delivers exceptional compute speed, affordability, and energy efficiency at scale. Groq solutions are based on the Language Processing Unit (LPU), a new category of processor.

Definition: AI specialized in performing a single task or a narrow set of tasks exceptionally well.

Example: Virtual assistants like Siri, recommendation engines on Netflix, or facial recognition software.

Status: ANI is everywhere today, powering most AI applications.

Definition: The definition of AGI is constantly debated, but in general it's AI that can perform any intellectual task a human can, demonstrating the ability to rationalize and generalize knowledge.

Example: An AGI would be able to code, calculate, play chess, interact, create new ideas, write a novel and so on (that's why some argue we're already there).

Status: AI has already passed Turing tests and is generally in par in many human domains, but it's not on par on all domains (yet).

My take: Currently AI is currently capable to perform most junior level intelligence jobs, but not able to replace specialized (senior level) developers. Once it's capable to surpass senior levels I would call it a AGI

Definition: AI that surpasses human intelligence across all domains.

Potential: ASI could solve complex global challenges like are we living in simulation, curing diseases or climate change, but it also raises significant concerns, because well, it's a LOT smarter than any of us.

Status: AI is currently capable to surpass us in narrow domains, and once it's capable to perform better in all intelligence related domains, it can be called ASI.

What It Does: Allows AI to learn and improve from experience.

Example: Predicting stock prices based on historical data.

Techniques:

• Supervised Learning: Training with labeled data (e.g., spam vs. non-spam emails).
• Unsupervised Learning: Finding hidden patterns in unlabeled data (e.g., customer segmentation).
• Reinforcement Learning: Learning through trial and error (e.g., AI mastering chess by playing itself).

What It Does: Mimics the structure of the human brain using artificial neural networks to process complex data.

Example: Recognizing objects in photos, translating languages in real time.

What It Does: Enables machines to understand and respond to human language.

Example: Chatbots, voice assistants, and language translation tools.

What It Does: Allows AI to analyze and interpret visual information.

Example: Autonomous vehicles detecting road signs, facial recognition.

What It Does: Creates new content (e.g., text, images, music) based on learned patterns.

Example: ChatGPT for text generation or DALL-E for image creation.

Units of text AI processes (e.g., words or parts of words).

• Example: The sentence “AI is fun” might be split into tokens like [“AI”, “is”, “fun”].

Internal settings AI adjusts during training to improve accuracy.

Numerical values that guide AI in making decisions.

• Analogy: Think of weights as the neurons in a human brain adjusting to learn new skills.

AI’s ability to keep track of context within a session (limited by token capacity).