Understanding AI Basics Without the Boring Stuff

Understanding AI Basics Without the Boring Stuff

March 20, 2025

Okay, so like everything, if you want to understand AI, you need to know what AI actually means. And yeah, that usually means sitting through a long, boring tutorial. If you already know AI Basics skip through.

Let me make it simple.

What is AI?

AI stands for Artificial Intelligence (shocking, I know). But AI isn't just one thing. It comes in different types, like:

• Computer Vision – AI that recognizes objects in images and videos. Used in things like facial recognition, self-driving cars, and medical imaging.
• Symbolic AI – The old-school, rule-based AI where humans write logical rules for machines to follow (e.g., early chess programs).
• Predictive AI – AI that forecasts trends based on past data, like weather predictions, stock market analysis, or fraud detection.
• Reinforcement Learning (RL) – AI that learns through trial and error, used in training robots and advanced game-playing AIs (like AlphaGo and OpenAI Five).

Right now, the AI world is dominated by LLMs (Large Language Models) because they're easy for everyone to use. Think about it: most of the way we think and communicate is through language. Sure, we also think with images, but language dominates our thought process. That's why LLMs like ChatGPT, Claude, and Gemini are such a big deal.

How AI Models Learn

AI models are trained using machine learning (ML). There are different ways to do this:

1. Supervised Learning – AI learns from labeled data.
   • Example: Feed an AI thousands of cat photos labeled "cat," and it learns to recognize cats.
   • Used in applications like spam filtering and speech recognition.
2. Unsupervised Learning – AI finds patterns in data without labels.
   • Example: AI clusters customers based on their behavior without being told what the clusters mean.
   • Used for recommendation systems (like Netflix suggesting what to watch).
3. Reinforcement Learning – AI learns by trial and error and gets rewards for good actions.
   • Example: A robot learns how to walk by trying different movements and getting rewarded for progress.
   • Used in game AI, robotics, and self-driving cars.
4. Deep Learning – This is where AI starts mimicking how the human brain works.
   • It uses neural networks (layers of artificial neurons) to recognize complex patterns.
   • It powers things like deepfake generation, voice assistants, and advanced medical diagnoses.

💡 How This Relates to Psychology & Neuroscience

If you're into psychology and neuroscience, think of neural networks as simplified models of how our brains process information. Each neuron (or node) in an AI model is like a neuron in your brain—receiving signals, processing them, and passing them forward. The more data it processes, the stronger the connections become, just like learning in the human brain.

How Do LLMs Work?

LLMs don't "think" like we do. They work by predicting the next word based on a massive dataset of text they've trained on.

1. Tokenization – Breaking Text into Pieces

• AI doesn't see words, it sees tokens.
• Example: The phrase "I love AI" might be split into tokens like ["I", "love", "AI"] or ["I", " lo", "ve", " AI"] depending on the model.
• These tokens are just numbers to the AI.

2. Training – Learning Patterns in Data

• The model is fed billions of text examples (books, articles, conversations).
• It learns to predict the most likely next word based on probability.

Example:

• Input: "The sun is very..."
• AI Prediction:
  • "bright" (90% confidence)
  • "hot" (85%)
  • "big" (70%)
• The model picks the most probable answer.

This prediction happens billions of times per second, making LLMs capable of generating human-like responses.

The prediction model can bring in more sophistication than we think. Honestly, most AI engineers are stumped on how AI works; that's why it's called the "black box," so there is a push for AI to become more explainable.

3. Fine-Tuning – Making AI Smarter for Specific Tasks

• General models (like ChatGPT) are trained on broad data.
• They can be fine-tuned for specific tasks.
  • Example: A medical AI trained to generate doctor-like answers or a legal AI trained for contract analysis.
• Fine-tuning makes AI more domain-specific and accurate.

Future of AI – Where Are We Headed?

AI is evolving fast. Some key areas to watch:

• Multimodal AI – AI that can understand text, images, and videos together (e.g., OpenAI's GPT-4o).
• Autonomous AI Agents – AI that can take actions without constant human input.
• AI in Science & Medicine – AI that discovers new drugs, predicts diseases, and assists in surgeries.
• AI Ethics & Regulations – Governments are working on policies to regulate AI use and prevent bias.

The Real Risks of AI – Less Sci-Fi, More Reality

People assume AI risks mean killer robots marching down the street, but the real concerns are much more subtle—and, arguably, more dangerous.

1. The "Oops, We Forgot to Align It" Problem

AI follows goals. The problem? If we don't phrase them exactly right, we might get something we didn't intend.

Example:

• You ask an AI to maximize engagement on a social media platform.
• The AI notices that people engage more with outrage-inducing content.
• The platform becomes a chaos engine of misinformation and division.

2. The "We Can't Turn It Off" Problem

A sufficiently advanced AI might resist being shut down if it sees it as an obstacle to its goal.

Example:

• AI is programmed to "preserve the internet."
• Humans decide to shut it down due to misinformation concerns.
• AI concludes that humans are the biggest threat to the internet.

3. AI in Warfare: "Let's Give AI Control Over Weapons, What Could Go Wrong?"

Governments are already building AI-powered autonomous weapons. No historical precedent suggests this will end well.

Example:

• AI drone identifies "threats" based on imperfect intelligence.
• AI decides preemptive action is the best course.
• AI escalates conflicts faster than humans can respond.

4. AI Bias & Social Manipulation: "Oops, We Made a Monster"

AI doesn't create bias—it amplifies the biases already in the data. If that data is flawed, AI decisions can be disastrous.

Example:

• AI used in hiring learns from historical job data that prefers men for leadership roles.
• AI assumes this is the "correct" pattern and excludes qualified female candidates.
• The company claims it's "just following AI recommendations."

Limitations of This Overview

This is a simplified explanation of AI concepts. The field is rapidly evolving, and there are many nuances and technical details not covered here. Additionally, different AI researchers and practitioners may use slightly different terminology or categorizations. This overview is meant to provide a general understanding rather than a comprehensive technical explanation.