The AI hierarchy, in order

To understand AI, visualize concentric circles. The outermost circle is Artificial Intelligence—the grand vision. Inside that is Machine Learning—the technique of learning from data. Inside that is Deep Learning—using neural networks. And at the cutting edge, we find Generative AI—models that create.

We classify AI capability into three stages:

• ANI (Artificial Narrow Intelligence): AI that excels at one specific task (e.g., playing Chess, recommending movies). This is where we are today.
• AGI (Artificial General Intelligence): AI that possesses the ability to understand, learn, and apply knowledge across a wide variety of tasks, matching human capability.
• ASI (Artificial Super Intelligence): An intellect that is much smarter than the best human brains in practically every field.

It is the convergence of three factors:

1. Big Data: The internet provided the fuel.
2. Compute Power: GPUs provided the engine.
3. Better Algorithms: Transformers provided the map.

• 1950: Alan Turing proposes the Turing Test.
• 1956: The term "Artificial Intelligence" is coined at Dartmouth.
• 1997: Deep Blue beats Garry Kasparov at Chess.
• 2012: AlexNet revolutionizes computer vision (Deep Learning boom).
• 2017: The "Attention Is All You Need" paper introduces Transformers.
• 2022: ChatGPT is released, bringing Generative AI to the masses.

Think of them as concentric circles. Artificial Intelligence is the broad discipline. Machine Learning is a subset within AI that allows machines to learn from data. Inside Machine Learning is Deep Learning, and within that are Artificial Neural Networks. AI processes data to make decisions; ML algorithms allow AI to learn and get smarter without additional programming.

Artificial Neural Networks mimic the biological brain, with nodes (neurons) grouped in layers working in parallel. They strengthen connections to improve pattern recognition.

Deep Learning involves many layers of these networks and huge volumes of complex data. It extracts features hierarchically: a system might recognize a plant in the first layer, a flower in the next, and a yellow daisy in the last.

1. Supervised Learning: Learning by example. The machine is given labeled inputs and outputs (e.g., images of daisies labeled "daisy"). It learns to map new inputs to the correct output.
2. Unsupervised Learning: No answer key. The machine analyzes unlabeled data to find hidden patterns, clusters, or structures, similar to how humans observe and categorize the world.
3. Semi-Supervised Learning: Uses a small amount of labeled data to guide the analysis of a large amount of unlabeled data. This speeds up learning and improves accuracy.
4. Reinforcement Learning: Learning by trial and error. The system receives "rewards" for good actions and "penalties" for bad ones, optimizing its strategy over time (e.g., playing chess).

• Recommendation Engines: Streaming services (Netflix, Spotify) analyzing viewing habits to suggest content.
• Dynamic Marketing: Analyzing customer data to personalize marketing and engage in real-time.
• ERP & Automation: Optimizing workflows and automating repetitive tasks using business data.
• Predictive Maintenance: IoT sensors on machinery predicting failures before they happen, saving costs and preventing downtime.

Bias and Spurious Correlations: Models can learn incorrect associations (e.g., correlating margarine consumption with divorce rates) if the data is flawed or if they find coincidental patterns.

The Black Box Problem: Complex models like deep neural networks can be difficult to interpret. It is often unclear how or why a specific decision was made, which poses risks in critical fields.

The "Deep" in Deep Learning refers to the number of layers in the neural network. Traditional neural networks might have 2-3 layers. Deep learning models can have hundreds or thousands. This depth allows the model to learn a hierarchy of features—from simple edges and textures to complex shapes and objects.

In traditional ML, humans had to manually select features (e.g., "does this image have ears?"). In Deep Learning, the network performs automatic feature extraction. It learns what features are important directly from the raw pixels or text.

Deep Learning is the technology behind self-driving cars, voice assistants, facial recognition, and the recent generative AI boom. It thrives on scale—more data and more compute usually lead to better performance.

• CNNs (Convolutional Neural Networks): The kings of computer vision.
• RNNs (Recurrent Neural Networks): Good for time-series and sequential data.
• Transformers: The state-of-the-art for natural language processing (NLP).

A neuron takes multiple inputs, multiplies them by weights (importance), adds a bias (threshold), and passes the result through an activation function (non-linearity). If the signal is strong enough, the neuron "fires" and passes information to the next layer.

• Feedforward NN: The simplest type. Information moves in one direction.
• CNN (Convolutional Neural Network): Specialized for processing grid-like data (images). It scans the image with filters to detect patterns.
• RNN (Recurrent Neural Network): Designed for sequential data (time series, text). It has a "memory" of previous inputs.
• Transformer: The modern architecture for language. It uses "attention" mechanisms to weigh the importance of different parts of the input data simultaneously.

They learn through a process called Backpropagation. The network makes a guess, compares it to the actual answer to calculate the loss (error), and then works backward to adjust the weights to minimize that error. This is repeated millions of times.

It must be broadly capable. Unlike previous models designed for one task (e.g., sentiment analysis), a foundation model can write poetry, debug code, translate languages, and summarize text, all without specific retraining.

Foundation models exhibit emergence—capabilities that were not explicitly trained for. For example, a model trained simply to predict the next word in a sentence might emerge with the ability to translate languages, write code, or solve logic puzzles.

The lifecycle involves two stages:

1. Pre-training: The expensive, compute-intensive phase where the model learns general patterns from massive datasets (e.g., "learning to read and write").
2. Fine-tuning: The adaptation phase where the model is specialized for a specific task or behavior (e.g., "learning to be a helpful assistant").

Prominent foundation models include the GPT series (OpenAI), BERT (Google), Claude (Anthropic), and Stable Diffusion (Stability AI).

Traditional AI is Discriminative: It draws a line to separate data (e.g., "Is this a cat or a dog?"). Generative AI is Creative: It learns the distribution of the data to create new examples (e.g., "Draw me a cat that never existed").

Generative models, such as Diffusion Models (for images) or Transformers (for text), learn the underlying structure of the training data. They then use probability to assemble new patterns that follow those structures but are not identical copies.

Generative AI is democratizing creativity. It acts as a co-pilot for writers, artists, coders, and designers, allowing them to iterate faster and explore new ideas. It is shifting the bottleneck from "skill" to "imagination".

• Marketing: Generating ad copy and visuals.
• Coding: Writing boilerplate code and documentation.
• Entertainment: Creating game assets and scripts.
• Science: Generating novel protein structures for drug discovery.

1. OpenAI: The pioneer. GPT-4 and ChatGPT set the standard for modern generative AI. OpenAI focuses on pushing the boundaries of scale and reasoning capability.
2. Google (DeepMind): The sleeping giant. With Gemini, Google has integrated its vast research capabilities into a multimodal model that integrates deeply with the Google ecosystem.
3. Anthropic: The safety-first contender. Founded by former OpenAI employees, Anthropic focuses on "Constitutional AI" and safety. Their Claude models are known for their large context windows and nuanced writing.
4. Meta (Facebook): The open-source champion. Meta's LLaMA series has been pivotal in enabling the open-source community to build and run powerful models on their own hardware.
5. Mistral: The European challenger. Based in France, Mistral produces highly efficient, open-weight models that rival the giants in performance-per-parameter.

• Closed Source (Proprietary): Models like GPT-4 and Gemini. You access them via API. They are generally more powerful and easier to use, but you have less control and privacy.
• Open Source (Open Weights): Models like LLaMA and Mistral. You can download and run them yourself. They offer privacy, control, and customization, but require hardware to run.

• Text & Writing: Tools like ChatGPT, Claude, and Jasper assist with drafting, editing, summarizing, and brainstorming text.
• Image & Design: Tools like Midjourney, DALL-E 3, and Stable Diffusion allow users to generate photorealistic images and art from text descriptions.
• Coding: Tools like GitHub Copilot, Cursor, and Replit act as pair programmers, suggesting code, debugging, and refactoring.
• Productivity: Tools like Perplexity (search), Otter.ai (transcription), and Notion AI (workspace) integrate AI into daily workflows to save time.

1. Define your goal: Are you writing, coding, or designing?
2. Check the model: What underlying model does it use? (e.g., GPT-4 vs Claude 3)
3. Consider privacy: Does the tool train on your data?
4. Look for integration: Does it fit into your existing workflow?

• ChatGPT (OpenAI): The industry standard for conversational AI.
• Claude (Anthropic): Known for safety and large context handling.
• Gemini (Google): Multimodal assistant integrated with Google apps.
• Perplexity: AI-powered search engine for accurate answers.

• Midjourney: Highest quality artistic image generation.
• Leonardo.ai: versatile asset generation for games and design.
• Runway: Leading tool for AI video generation and editing.

• Cursor: The AI-first code editor.
• GitHub Copilot: The most widely used code completion tool.
• V0.dev: Generative UI system by Vercel.

AI concepts like "Gradient Descent" or "Backpropagation" can be abstract and mathematical. Interactive visualizations allow you to see the math in action, building a deeper, intuitive understanding.

• Neuron Visualizer: See how inputs, weights, and biases combine to fire a neuron.
• Network Playground: Build and train simple neural networks in your browser.
• Gradient Descent Sim: Visualize how models minimize error by descending a loss landscape.
• Hyperparameter Sandbox: Experiment with learning rates and batch sizes to see their effect on training.

We can categorize intelligence into three distinct forms that interact in the modern world:

Definition: The ability of machines to think and reflect like humans, attempting to replicate human intelligence with machines.

Capabilities:

• Reasoning: Logical thinking and inference.
• Natural Communication: Understanding and generating human language.
• Problem-solving: Finding solutions to complex challenges.

Characteristics:

• Replaces Human Effort: Automates tasks traditionally done by humans.
• Performs Tasks Independently: Operates without constant human intervention.

"Why aren't we more compassionate?" — Daniel Goleman

Emotional Intelligence operates through the basal ganglia (the wisdom center of the brain). It guides decisions based on emotional valence (what felt good/bad). Unlike the neocortex, it does not speak in words but is connected to emotions and the gut.

Key Insight: Combine EQ + IQ for better decisions.

Algorithm: A set of rules or instructions given to an AI, neural network, or other machine to help it learn on its own.

Alignment: The problem of ensuring AI systems have goals that match human values.

Bias: Errors in AI output resulting from prejudices in the training data.

Fine-tuning: The process of training a pre-trained model on a smaller, specific dataset to specialize it.

Hallucination: When an AI generates false or nonsensical information confidently.

LLM (Large Language Model): A deep learning algorithm that can recognize, summarize, translate, predict, and generate text.

Multimodal: AI that can understand and generate multiple types of media (text, images, audio).

Parameters: The internal variables (weights) that the model adjusts during training. GPT-4 has trillions.

Token: The basic unit of text for an LLM (roughly 0.75 words).