How do you feel about AI? Like it? Love it? (Some people literally do, more than you might think.) Hate it? What do you think it is? What do you think it isn't? Chances are you are probably partially right, partially wrong, and have highlighted numerous unknowns. And that isn't your fault: few inventions or technologies in human history have been so many different things simultaneously, deployed globally this quickly, with as many benefits and risks, as AI.
But when we talk about AI, are we all having the same conversation? (Spoiler alert: we aren't.) Should LLMs (Large Language Models) like ChatGPT and Claude and Perplexity be spoken about in the same breath as algorithms that steer social media content our way, or that processed medical data in the development of COVID-19 vaccines, or that are making computer coding roles obsolete? There's the age-old expression 'not seeing the forest for the trees.' If Grok or Claude or ChatGPT or this or that are the trees… what is the forest?
PART 0: WHAT YOU ALREADY (PROBABLY) KNOW
ChatGPT, Gemini, Claude, Llama, Nova Micro (Amazon's model), Perplexity, Copilot, Grok. Email composition, photo editing, video generation, search result summaries. Your Spotify DJ. Finding AI in daily life nowadays is not hard; in fact, it's probably harder to avoid it. These are just the consumer-facing applications and services actively marketed to us—tools companies advertise directly with promises of efficiency, productivity, and insight.
But LLMs and countless apps are the tip of the iceberg. Companies (and governments) have tools and capabilities you probably don't know about or haven't conceived of. When was the last time you noticed a new API integration in your favorite app? This isn't surprising: tools, services, and browser extensions are multiplying rapidly. So let's slow things down and start at the beginning.
PART 1: AI 101 — WHAT IT IS, HOW IT WORKS, AND WHY IT MATTERS
Key Terms Used in This Guide
Before we dive in, here are essential terms you'll encounter:
Training: The process of exposing an AI system to massive amounts of data so it can learn patterns
Model: The actual AI system after training—a specific configuration of millions or billions of parameters
Parameters: The numerical values inside a neural network that determine how it transforms input into output
Inference: Using a trained AI model to generate outputs (predictions, text, images, etc.)
Neural Network: A mathematical structure loosely inspired by biological brains, consisting of layers of interconnected nodes
LLM (Large Language Model): An AI system specifically trained on text to predict and generate language
From Science Fiction to Everyday Reality
For decades, artificial intelligence belonged almost exclusively to science fiction—appearing in novels and films as either existential threat or revolutionary assistant. AI was more concept than capability—until suddenly, it wasn't. Then, over the past decade, AI systems gradually slipped into the fabric of daily life. There was no single announcement or public launch day. AI simply became increasingly accessible through cloud services and consumer applications. Anyone with a smartphone or laptop could suddenly interact with tools that felt conversational and surprisingly capable.
This sudden accessibility fundamentally shifted public perception and expectations. As of September of last year, more than 60% of American adults use it regularly, and globally, nearly 2 billion, of Earth’s 8.2 billion people, use it. Yet the pace of AI's expansion has left many people uncertain about what it actually is. Understanding AI is no longer a futuristic curiosity; it is part of understanding modern life.
What AI Really Is
The phrase "artificial intelligence" suggests machines that think and perceive like humans. Despite the name, that's not what current AI systems are. To be absolutely clear: today's AI systems do not think like humans. They do not think at all. Modern AI systems, including LLMs, do not possess awareness, self-reflection, or genuine understanding by current scientific understanding. They do not form intentions, experience understanding, or possess goals.
Instead, what is commonly called 'AI' is best described as sophisticated pattern recognition and statistical prediction. These systems learn by analyzing vast quantities of data and identifying regularities in language, images, audio, code, or sensor signals. Given an input, they predict the most likely output—such as the next word in a sentence or the label for an image—based on learned statistical patterns.
The mathematical engine behind most modern AI is the neural network. Despite the biological-sounding name, neural networks are nothing like actual brains. They're mathematical structures consisting of layers of nodes (called 'neurons'), each connected to nodes in adjacent layers. Here's how they work:
Data enters the first layer as numbers (words converted to numbers, pixels as numbers, etc.)
Each node performs a simple mathematical operation: it multiplies the inputs by learned weights, adds them together, and passes the result through a simple function
Results flow through multiple layers, with each layer performing similar transformations
Final layer produces output (a predicted next word, an image classification, etc.)
The 'learning' happens when the system adjusts those weights (called parameters—billions of them in large models) to minimize the difference between its predictions and the correct answers in training data. This process, called 'training,' requires enormous computational power and massive datasets, but the underlying principle is just calculus-based optimization.
The system doesn't make 'decisions' or 'assumptions.' It transforms numerical inputs into numerical outputs through millions of mathematical operations. The results seem intelligent because, with enough parameters and enough training data, these mathematical transformations can encode remarkably complex patterns.
Think of AI as an extraordinarily sophisticated calculator that has processed billions of examples. It doesn't understand concepts in any meaningful sense, but it produces outputs that appear to demonstrate understanding because it has encountered countless similar patterns in training. The power lies in scale and probability, not in insight or consciousness.
There's ongoing debate about whether 'pattern recognition' adequately describes what these systems do. Some researchers characterize them as 'stochastic parrots'—systems that produce statistically plausible text without any form of understanding. Others point to surprising emergent capabilities in large models as potential evidence of something more sophisticated than pure statistical correlation. For our purposes, what matters is that whatever these systems are doing internally, they demonstrably lack human-like comprehension, consciousness, or intentionality.
How AI Learns Without Comprehending
The core learning process is conceptually simple but requires staggering computational resources. Training a model involves feeding it millions or billions of examples. These can include:
Labeled data: Examples with correct answers (e.g., images tagged as "cat" or "dog")
Weakly labeled data: Examples with approximate or partial labels
Unlabeled data: Raw text, images, or audio where the model learns patterns without explicit answers
The model processes these examples—text documents, images, code, audio—and adjusts billions of internal numerical parameters to minimize prediction errors. Through repeated exposure, the model encodes intricate mathematical relationships between inputs and outputs in billions of numerical parameters.
Some models are trained once and then frozen. Others are fine-tuned on narrower data or shaped with human feedback. A common approach, reinforcement learning from human feedback (RLHF), has humans rank or rate model responses; the model is then updated to prefer outputs that align better with those rankings. RLHF does not grant human judgment, but it nudges the model toward responses that feel more helpful, safe, or polite.
However, RLHF has limitations. It can make models refuse harmless requests if human raters were overly cautious, and it may not fully align models with any specific individual's values, since 'human feedback' represents an aggregation of diverse—and sometimes contradicting—preferences from many different raters.
The result is a system that can generate plausible text, classify images, or interpret speech, often giving the impression of understanding while still lacking human-like comprehension.
KEY TAKEAWAY: AI learns through exposure to massive datasets, adjusting billions of internal parameters to minimize prediction errors. It's not "teaching" in the human sense—it's statistical optimization. The model doesn't acquire understanding; it acquires the ability to produce statistically likely outputs that often appear to demonstrate understanding.
Why AI Feels Different From Traditional Software
Traditional software is rule-based and brittle. It follows explicitly coded instructions and often fails when inputs deviate from what the programmer anticipated. A missing field, a wrong data type, or a slightly different input format can cause errors or crashes.
AI systems, especially LLM-based tools, behave differently. They can tolerate ambiguity, incomplete instructions, or imprecise language and still attempt a response. They may ask clarifying questions, infer missing details, or provide their best guess. That flexibility creates an impression of collaboration and makes interactions feel closer to conversation than to command syntax.
This shift—from programming computers step by step to describing goals in natural language—is one of the most important changes in modern computing.
Real-World Impact That Is Already Here
AI is already embedded in critical systems around the world. In specific controlled studies, AI has matched or exceeded human performance at certain medical image analysis tasks, such as detecting diabetic retinopathy or identifying suspicious lesions in mammograms. These tools support clinicians rather than replace them, offering a second set of eyes in high-volume or high-stress environments.
However, medical AI presents a complex picture. While some studies show AI matching or exceeding human radiologists at specific tasks, other research reveals significant limitations:
AI trained on datasets from wealthy countries performs poorly in resource-limited settings where imaging equipment differs
Models can 'cheat' by learning to identify the specific scanner or hospital rather than actual medical conditions
A system that appears 95% accurate overall might be 99% accurate for one demographic and 85% for another, with these disparities hidden in averaged statistics
The key insight: AI in healthcare is neither miracle cure nor total failure—it's a powerful tool whose benefits depend entirely on thoughtful deployment, diverse training data, continuous monitoring, and keeping expert clinicians in the decision-making loop.
Beyond medicine, AI supports fraud detection in finance, supply-chain optimization in logistics, predictive maintenance in manufacturing, crop and weather forecasting in agriculture, and language translation across global communication. Everyday features such as spam filters, autocomplete, search ranking, and social feed curation are powered by narrow AI systems working quietly in the background.
However, the deployment of AI in these critical systems raises important questions. In healthcare, while AI can spot patterns doctors miss, it can also inherit biases from training data—for example, skin condition detection systems trained predominantly on light-skinned patients may perform poorly on darker skin tones. Financial AI might deny loans to qualified applicants based on patterns in historical data that reflect past discrimination. These systems can appear objective while actually encoding and amplifying existing inequalities. The key question isn't whether AI can do these tasks, but whether its deployment in high-stakes domains is being done thoughtfully, with appropriate human oversight and continuous monitoring for unintended harms.
AI's influence is broad and often invisible, but it shapes decisions, experiences, and outcomes in concrete ways.
Strengths and Limitations
AI's strengths are clear. It excels at processing large datasets quickly, spotting patterns that humans might miss, and delivering personalized experiences at scale. It enables individuals, small teams, and smaller organizations to access capabilities—like instant translation or code generation—that once required specialized expertise.
Its limitations are less immediately obvious—masked by sophisticated language and stunning speed—but it's important that we slow down and 'chew our food carefully' so to speak. If we're using AI for some purpose and aren't able to distinguish when we're in its blind spot, we can weave confusion, error, and false assumptions into our impression of reality and truth—which can be costly, consequential, even dangerous.
Current models don't have an inherent, foundational concept of truth. They weren't designed to verify facts against reality or maintain internal consistency. Instead, they predict likely text based on training data. However, because they're trained largely on human-created text (which generally aims to be truthful) and optimized with human feedback (which rewards accuracy), they often produce factually correct outputs—not because they 'know' what's true, but because true statements were prevalent in training data and rated highly by human reviewers.
This is why AI can be simultaneously impressive and unreliable. When the answer is well-represented in training data and human raters agreed on it, AI performs excellently. When questions involve recent events, specialized knowledge, or subtle context, accuracy drops significantly—and the system has no internal signal telling it which answers to trust.
This point is extremely important to keep in mind: interacting with AI might feel like a fact-based exchange, but one should always take an AI response with a grain of analog salt. This can lead to "hallucinations": confident, fluent statements that are simply wrong. Models also learn from historical data and can inherit or amplify biases present in that data, especially around sensitive topics such as race, gender, or geography.
More fundamentally, current AI systems face several structural limitations:
No persistent memory or true learning: Unlike humans who accumulate knowledge over time, most AI systems cannot update their knowledge after training. They cannot learn from conversations or remember you between sessions (unless explicitly programmed with memory features).
Brittle generalization: AI can perform brilliantly on tasks similar to its training data but fail completely on slight variations. A model that excels at detecting cats in photos might fail if shown a cat wearing a costume, not because it's 'harder' but because it's unexpected.
Energy and resource intensity: Training large AI models consumes enormous amounts of electricity. Training GPT-3 once generated an estimated 552 metric tonnes of CO2 equivalent—roughly the lifetime emissions of five average American cars. Inference (using the model) also requires significant computational resources.
Opacity and explainability: When an AI system makes a decision—denying a loan, flagging a résumé, recommending a medical treatment—it's often impossible to understand exactly why. The decision emerges from billions of parameter interactions, creating 'black box' systems that raise accountability concerns in high-stakes applications.
They struggle with tasks that require deep causal reasoning, long-term planning, or ethically charged judgment without strong human oversight. In high-stakes environments—medicine, law, critical infrastructure—over-reliance on AI without robust review can create real risk.
Recognizing both strengths and weaknesses helps frame AI as a powerful tool that still requires human responsibility, context, and critical thinking.
What This Means for You
Understanding that AI predicts rather than comprehends has practical implications:
Don't trust AI for critical factual claims without verification - it optimizes for plausibility, not truth
AI works best for tasks with clear patterns - writing assistance, summarization, pattern recognition
Human judgment remains essential for context, ethics, and high-stakes decisions
Ask: "Where would this AI's training data come from?" to identify potential blind spots
Common Misconceptions About AI
Myth: "AI is intelligent like humans are intelligent"
Reality: AI performs impressive pattern matching without comprehension, consciousness, or understanding
Myth: "AI is always objective and unbiased"
Reality: AI inherits biases from training data and human feedback
Myth: "AI will replace most human jobs soon"
Reality: AI augments human capabilities for some tasks but lacks the flexibility, common sense, and contextual understanding for most complete job replacement
Myth: "AI can't create anything original"
Reality: AI generates novel combinations and variations of patterns from training data—whether that counts as "original" is philosophical
Myth: "AI development is unstoppable"
Reality: AI development responds to funding, regulation, public pressure, and technical limitations—its trajectory is not predetermined
Why Understanding AI Matters
Understanding AI is not just another form of digital literacy—it's becoming essential to navigating modern life. As AI systems increasingly mediate access to information, employment, credit, healthcare, and legal services, knowing their capabilities and limitations becomes crucial to self-advocacy and informed citizenship.
This understanding doesn't require expertise in machine learning mathematics. Rather, it means developing the conceptual clarity to ask: Where does this AI's training data come from? Who decided what counts as a 'good' output? What happens when the system encounters something unexpected? Who benefits from this application, and who might be harmed?
AI is not just another software upgrade. It represents a structural shift in how information is created, transformed, and acted upon across communication, policy, creative work, healthcare, education, and more. As AI systems become more tightly woven into institutions and daily routines, understanding how they work becomes a form of digital and civic literacy.
This does not require mastering advanced math or building models from scratch. It means having enough conceptual clarity to ask good questions, evaluate where AI fits in a decision, and recognize when human judgment must remain in charge.
The Deeper Questions AI Raises
Beyond practical applications, AI forces us to confront profound questions:
What is intelligence? If a system can perform tasks we consider intellectual—writing, reasoning, problem-solving—without understanding, consciousness, or intent, what does that tell us about intelligence itself? Perhaps many tasks we assumed required comprehension are actually pattern-matching at scale.
What is authentic? As AI-generated text, images, and code become indistinguishable from human-created work, how do we value human creativity? Does it matter whether a poem, painting, or essay was created by a person if it moves us equally?
What should remain human? Some argue certain decisions—medical diagnoses, legal rulings, hiring decisions—should always involve human judgment, even if AI could perform them more efficiently. Others contend that human judgment is often biased and that AI (properly designed) could be more fair. Where should we draw these lines?
Who controls the technology? A handful of corporations currently control the most powerful AI systems. Their choices about training data, safety measures, and deployment shape outcomes for billions of people. Should AI development be more democratic, transparent, or regulated? Or would that stifle innovation and cede leadership to less scrupulous actors?
These questions don't have clean answers—and your position on them will shape how you view AI's role in society.
A Critical Perspective on Technology Narratives
Throughout discussions of AI (and in the blockchain articles you've written), there's often an implicit "trickle-down technology" narrative: innovations adopted by institutions eventually benefit everyday people. This framing deserves scrutiny.
The efficiency gains often don't trickle down as promised. When brokerages became more efficient through technology, did average investors see proportional benefits, or did institutions capture most gains? High-frequency trading made markets more 'efficient' but arguably increased volatility and favored those with capital for advanced infrastructure.
Technology can consolidate power. AI tools might increase inequality by giving sophisticated users massive advantages while making others dependent on black-box systems they can't audit or understand. Similarly, blockchain adoption by financial institutions might optimize their operations without meaningfully improving consumer experience or access.
The "inevitable adoption" framing removes agency. Saying blockchain or AI "is coming whether you like it or not" discourages critical evaluation and public input into how these technologies are deployed. Whether technological adoption is 'inevitable' or 'democratizing' depends heavily on policy choices, regulatory frameworks, and public pressure—not just technological capability.
This isn't to say these technologies lack benefits—but rather that we should examine who benefits most, what trade-offs exist, and whether alternative implementations might serve broader interests.
Sources and Further Reading
With this foundation established, we can now examine how these AI systems form a modular ecosystem, specialized tools that can be combined, mixed, and matched to create increasingly powerful capabilities. But that's a topic for another piece.