This section was designed for readers who would like background on AI basics.  It consists of three five minute lessons:

Lesson 1 – What is AI?

Lesson 2 – Machine learning

Lesson 3 – Chat GPT

Lesson 1 – What is AI?

“AI is one of the most important things humanity is working on. It is more profound than electricity or fire.” – Sundar Pichai (CEO Google)

The term “artificial intelligence” (AI) was introduced at a 1956 workshop organized by Dartmouth mathematician John McCarthy.  According to Melanie Mitchell’s excellent introduction to the field (Artificial intelligence: A guide for thinking humans, p. 19) “McCarthy later admitted that no one really liked the name… but ‘I had to call it something, so I called it Artificial Intelligence.’”

For the next several decades, AI was an obscure academic specialty which most people knew little about.  That all changed on November 30, 2022, when OpenAI released ChatGPT 3.5 (see Lesson 3).  This chatbot can draft an email announcing a divorce, list steps you’ll need to take to plan for a move, give you ideas for bedtime stories to make up for your children, write computer code, and much much more. 

Similar capabilities had begun to emerge in a number of programs released before then, but ChatGPT dramatically improved performance and was the first to make it easily accessible to the general public, for free.  It became one of the fastest growing software products ever.  In the five days after its release, over one million people tried ChatGPT.  This led to a tsunami of interest and investment in AI. 

AI has already made possible dozens of products that we use every day.  Google Search, Apple’s Siri, Amazon’s Alexa, recommendation systems from Netflix and Spotify, Microsoft’s CoPilot, Roomba vacuum cleaners, self-driving cars from Tesla and Waymo, and facial recognition systems to increase airport security all include AI.

Behind the scenes, AI robots now play a major role in US auto manufacturing including vehicle assembly, welding, painting and quality control.  In China, manufacturers have developed “dark factories”: fully automated facilities operated by robots without human workers.  One dark factory in Beijing currently “operates round-the-clock producing one smartphone per second – with zero humans employed.”  (For more, see these posts in my China blog:  the AI Race, and  the Robotics Race.)

While no one can predict the effect of AI on human workers 20 years from now, there can be no doubt that it is already changing the world of employment.  When researchers from the University of Maryland “analyzed 186,000 [newspaper] articles published online” in 2025, they found that “around 9% of [the] articles… were at least partially written by AI.”  As newspapers shrink staff, that percentage is sure to increase.

Despite its widespread impact in recent years, there are still many differences of opinion about exactly how to define artificial intelligence, and where to draw the line between programs which include AI, and those which do not.  In discussing this problem, Mitchell’s book begins (p. 19) by quoting Voltaire’s warning to “Define your terms … or we shall never understand one another.” She goes on to note that this is “a challenge for anyone talking about artificial intelligence, because its central notion—intelligence—remains so ill-defined.”  Similarly, in the New York Times bestseller Empire of AI (p. 91) Karen Hao pointed out that “throughout history, neuroscientists, biologists, and psychologists have all come up with varying explanations for what [intelligence] is.” 

The ambiguity of the term AI is a dream come true for marketers who want to take advantage of a hot market by using the term, whether their new product contains any features that could remotely be classified as AI or not.  But some experts believe that given the primitive state of the field, the lack of a single definition of AI is actually a good thing.  As an article in IEEE Intelligent Systems summed it up: “Because we don’t deeply understand intelligence or know how to produce general AI, rather than cutting off any avenues of exploration, to truly make progress we should embrace AI’s ‘anarchy of methods.’”  In other words, what matters is whether a program is useful, not whether it can be technically classified as AI.  

How many types of AI are there?  It depends on who you ask.  A number of taxonomies have been proposed based on a wide variety of criteria including a program’s capability, architecture, learning paradigm, application domain, and operational characteristics. 

Perhaps the most important AI distinction, and one which is often misunderstood, is between narrow AI and Artificial General Intelligence (AGI).  Narrow AI systems are designed to perform specific tasks such as detecting fraud, analyzing CAT scans, and all of the examples quoted above.  Every AI system that has been developed to date falls into this narrow category. 

ChatGPT, Alexa, and self-driving cars are absolutely amazing, but all are examples of narrow AI.  I don’t know anyone who would classify them as “more profound than fire.”  That statement refers to AGI.  According to an AI research center at Stanford University, AGI can be defined as “an AI system with general, human-level (or beyond) ability to learn, reason, and apply knowledge across a wide range of tasks and domains… The term is controversial [in part because]… there’s no universally accepted test, so claims are hard to verify.”

Science fiction writers have been speculating about the risks of AGI for decades and have recently been joined by scientists and academics.  Philosopher Nick Bostrom’s 2014 book Superintelligence: Paths, Dangers, Strategies imagined an AGI that was designed to produce as many paper clips as possible.  If the program determined that humans were interfering with this goal, it could choose to kill us all.  Not to mention its risks to privacy, security, fairness, and the jobs of all the people who work in paper clip factories. 

These are all genuine and serious concerns.  But it is important to remember that when people talk about AI risks, they are usually talking about the risks of AGI, not of programs that actually exist today. 

Of all the controversies regarding AGI, the biggest is when, or even if, it will exist.  In 2023, Elon Musk wrote on Twitter Spaces that it will appear “in 5 or 6 years.”  Check back in 2029 if you’d like to see how accurate he was. 

In a breathless 2024 blog post, OpenAI founder Sam Altman wrote that an “Intelligence Age characterized by ‘massive prosperity,’ would soon be upon us, with superintelligence perhaps arriving as soon as in ‘a few thousand days’… Although it will happen incrementally, astounding triumphs — fixing the climate, establishing a space colony, and [discovering new laws in] physics — will eventually become commonplace.” (Lao, p. 19)

As you consider these claims, it is helpful to remember the Danish proverb “it is very hard to predict, especially about the future.”  The entire field of AI is littered with optimistic AGI projections that have proven incorrect.  The most famous came from Nobel prize winning economist Hebert Simon in a 1965 book (p. 95) “machines will be capable, within twenty years, of doing any work a man can do.”  In case you dozed off somewhere around 1985, Simon was wrong. 

In addition, some experts argue just the opposite. In her book, Mitchell noted (p. 276) that “Several surveys given to AI practitioners, asking when general AI or ‘superintelligent’ AI will arrive, have exposed a wide spectrum of opinion, ranging from ‘in the next ten years’ to ‘never.’ In other words, we don’t have a clue.”

AI is currently at a Wild West stage in which researchers sometimes disagree about where to draw the line between programs that are or are not AI.  And if that’s not confusing enough, there’s the added complication that no one understands these programs work. (For more on this, see the discussion of black box models in Lesson 2.)

Putting it all together, “The field of AI is in turmoil” according to Mitchell (p. 13). Depending on whom you believe “either a huge amount of progress has been made, or almost none at all. Either we are within spitting distance of [AGI], or it is centuries away. AI will solve all our problems, [or] put us all out of a job, [and] destroy the human race.”

I guess we’ll find out.  But in my opinion, given that AGI does not exist now and may never exist, its risks rank near the bottom of my personal list of worries.  As Andrew Ng, an associate professor at Stanford, put it at the 2025 GPU Technology Conference “I don’t work on not turning AI evil today for the same reason I don’t worry about the problem of overpopulation on the planet Mars.”  The human race has far bigger problems to worry about in the foreseeable future, such as nuclear war, climate change, food and water shortages, political instability, economic inequality and pandemics.

The approach in this blog is therefore to spend little time talking about AI’s possible long-term risks.  Instead, it will focus on narrow AI products that are already changing the way professionals work, for better or for worse.

Lesson 2 – Machine learning

At this moment in history, machine learning is the most significant and most common approach to AI.  It can be broadly defined as an AI system that learns from data. According to Stanford’s Artificial Intelligence Index Report 2025, over 75% of recent AI research articles focus on machine learning. All of the familiar programs mentioned in Lesson 1 are based at least in part on machine learning. 

Many of the capabilities of these narrow AI programs are absolutely astounding.  Consider, for example, Netflix’s recommendation system. If you are a regular Netflix user, you have probably used their ever changing lists of personal recommendations to decide whether you should next watch My Litle Pony,  Lady Chatterley’s Lover, or another show based on your past viewing habits. 

To come up with these recommendations, AI analyzes every single program you have ever watched on Netflix.  Not to mention how you rated each program, whether you watched the whole show, turned it off midway, or watched some scenes over and over.  It also stores data on which recommended titles you have chosen over the years and which you ignored.

How large is the resulting database?   Netflix has 325 million subscribers in 190 countries around the world.  Each country has a slightly different list of its 10,000 plus shows available, depending on copyright laws and regional license agreements.  Multiply those numbers by the billions of times anyone in the world has used their remote to make a Netflix choice and the number of data points is ridiculously large.  And the chance is low that our little brains can really comprehend how trillions of variables interact to produce Netflix recommendations.

However, the basic principles behind machine learning systems are straightforward.  At the simplest level, virtually all computer programs can be conceptualized as follows:

Inputs  >>>  Processing  >>> Outputs

The first factor that makes machine learning different from traditional computer programs is the sheer number of inputs required.  Though a small machine learning program could be built to fit on a single PC, the AI programs we use in daily life generally require supercomputers in the cloud.  They can use millions of processors operating in parallel, so that impossibly large computational tasks can be broken down into millions of smaller pieces, all of which are operated on at the same time.

The second factor that distinguishes machine learning from other approaches is the complexity of the processing step.  This is where the magic happens. 

Traditionally, computers process data by applying a set of algorithms – steps and rules like the recipes in a cookbook.  For example, to count the number of items in a list, a programmer could set a counter at 0, go through each item in the list and add 1 to the counter.

But in machine learning, the computer creates its own algorithms, based on trial and error plus feedback.  It does this by means of neural networks and deep learning.  According to Karen Hao’s bestseller The Empire of AI (p 98): “At their core, neural networks are calculators of statistics that identify patterns in old data—text, pictures, or videos—and apply them to new data.”  For complex patterns such as spoken language, deep learning uses neural networks with many layers. 

In biology, neurons are of course the cells that communicate messages in the three pound organ that enable our species to build the Great Wall of China and remember where we left our keys. AI researchers borrowed the term neural network based on the highly controversial idea that artificial intelligence programs can serve as a model of how the human brain operates.  If you are the type of person that wants to know the details of how these complex systems work, I’d recommend that you start with the online lesson on neural networks at www.3blue1brown.com/lessons.  For everyone else, I avoid that rabbit hole and focus on what AI can do rather than how it does it. 

In machine learning, processing begins when a computer analyzes the input data, and looks for some pattern, any pattern.  It gradually figures out rules for itself by trial and error.  For example, imagine a program that is designed to look at a large assortment of photos, and identify which photos include a dog and which do not.  The computer randomly guesses on the first photo and is then given feedback on its answer.  The model is adjusted based on this feedback to create a second, more accurate, model.  Then the second model repeats the steps to create a third model, a fourth, a fifth, and so on.  This cycle repeats over and over in a massive feedback loop that could recur literally billions of times, until it reaches an acceptable level of accuracy, and is ready for use. 

There are three main types of feedback that are used to develop chatbots:   

1. In supervised learning, the program is trained from a dataset of examples which have been pre-labeled with the correct output – e.g. it either includes a dog, or it doesn’t.  This training database could include millions of photos, each labeled dog or no dog.  If the program identifies an image incorrectly, its model will be adjusted.  After all images have been rated, a new model will be created, and the process will be repeated.  Over and over, until the computer correctly answers not just the obvious examples – like a Saint Bernard napping in front of a fireplace – but also more subtle cases such as a picture of a parade, with a tiny image of one bystander holding a chihuahua.       
2. Unsupervised learning does not require a pre-labeled training dataset.  Instead, it bases its feedback on patterns of past behavior.  For example, to detect credit card fraud before it occurs, banks maintain huge databases of past legitimate vs fraudulent transactions.  An AI system then looks for patterns associated with fraud.  If it spots something suspicious, you may get a text that says: “Alert.  Did you try to use your VISA card from Loans R Us to spend $187.92 at the Walmart in Little Rock Arkansas?  Reply YES or NO. If NO, your card will be blocked.”
3. Reinforcement learning can be used to increase user engagement by rewarding users’ past choices.  For example, if YouTube recommends that you look at a funny video of a cat stalking and attacking a cat balloon, it’s based on what you’ve watched before.  Each time you click on a recommended video, the model is updated to reflect your preferences.  If you watch a whole video through, it will be rated more positively than one that you rapidly click away from or never choose in the first place.

There are also many other types of feedback used in AI training, some of which rely on human ratings of sample responses from the system being trained.

When a pattern detection model completes its analysis of hundreds of thousands of variables and their interactions, a reasonable observer might wonder: how is the computer doing that?  What rule is the model using when it decides that one picture includes a dog, but another does not?  The answer, according to AI researcher Joel Dudley is very simple: “We can build these models, but we don’t know how they work.”

The article that quoted Dudley was published in the MIT Technology Review and titled “The dark secret at the heart of AI… No one really knows how the most advanced algorithms do what they do.”  The article went on to say that “the interplay of calculations inside a deep neural network is crucial to higher-level pattern recognition and complex decision-making, but those calculations are a quagmire of mathematical functions and variables. 

Or, in the words of Karen Hao (p. 107) “pop open the hood of a deep learning model and inside are only highly abstracted daisy chains of numbers. This is what researchers mean when they call deep learning ‘a black box.’ They cannot explain exactly how the model will behave.”   

In some cases, this lack of transparency can be associated with significant errors.  For example, in her light-hearted AI introductory book You Look Like A Thing And I Love You (p. 25), Janelle Shane gave an example of an experiment she did years ago, on one of Microsoft’s first image recognition products, to test its ability to recognize sheep in a variety of environments.  “One day I noticed something odd about its results: it was tagging sheep in pictures that definitely did not contain any sheep. When I investigated further, I discovered that it tended to see sheep in landscapes that had lush green fields—whether or not the sheep were actually there.  The AI had been looking at the wrong thing. And sure enough, when I showed it examples of sheep that were not in lush green fields, it tended to get confused.”

Problems like this can be solved over time as programs become more sophisticated.  In any case, the sometimes weird errors made by AI programs are very useful for article headlines but are of little relevance to the typical user.

Nevertheless, the issue of lack of transparency lingers disturbingly in the background.  As the MIT Technology Review article summed it up: “We’ve never before built machines that operate in ways their creators don’t understand. How well can we expect to communicate—and get along with—intelligent machines that could be unpredictable and inscrutable?”

Lesson 3 – Chat GPT

The first time I used ChatGPT, it seemed absolutely magical. 

As researcher Kevin Roose put it in the New York Times, “I felt… like I was talking to something that could think.” Henry Kissinger was so impressed that he told an Atlantic reporter that “AI may be on the cusp of a new form of reason.”

When ChatGPT 3.5 was released by OpenAI in 2020, it quickly became the first commercially successful chatbot – a program that can converse with humans, using text or voice.  It was not the first program to simulate human conversation, but it was by far the most compelling. 

Encouraged by the public reaction, competitors raced to release chatbots of their own.  As of May 2026, five companies dominate 99% of the chatbot market:  OpenAI (ChatGPT has 60.6% market share with 800 million active users per week), Google (Gemini – 15.1% market share), Microsoft (CoPilot – 12.5%), Perplexity (the company and the chatbot share the same name – 5.4%) and Anthropic (Claude – 5.0%).  Today, these products are being applied in an ever-growing number of situations, ranging from drafting catchy marketing slogans to giving exercise advice, translating over 80 languages, writing computer code, and much more.

How can they possibly do all that?  All five of these chatbots are built on a foundation of large language models (LLMs), AI machine learning programs that have been trained to recognize language patterns derived from enormous amounts of data. 

As early as the 1980s, AI researchers experimented with statistical language models that could predict one word at a time based on the words that came before.  For example, given the phrase “I drank a cup of ___,” the model might fill in the blank with coffee or tea.  But it was only in 2017, when Google published a paper describing a new type of neural network called a transformer that statistical language models took off.  Transformers enabled AI programs to predict the next word in a sentence not just by looking at the preceding word, but also at the larger context of the sentences and paragraphs surrounding it. 

The next year OpenAI released GPT-1 (the first Generative Pre-Trained Transformer) which, according to Forbes, “demonstrated the power of unsupervised learning in language understanding tasks, using books as training data to predict the next word in a sentence.”  But it wasn’t that useful.  In her bestseller Empire of AI (p. 124), Karen Hao noted that “Compared with today’s models, the text produced [by GPT-1] was clunky and often descended into gibberish.”

In 2019, GPT-2 increased the number of parameters (numerical values that the program learns during training) more than 12 times (from 117 million to 1.5 billion) and the results were shocking… simply making the model bigger produced radically better performance. When OpenAI took the obvious next step and increased the number of parameters over 100 times more (to 175 billion) in the LLM GPT-3.5 Turbo, performance improved even more dramatically.  This version was so powerful, that in the words of the team that developed it, “GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans.”  Soon after, ChatGPT 3.5 was released to the public, and the rest is history.  

This is the most famous example suggesting that when it comes to AI, bigger is better. In the words of Dario Amodei, who founded Anthropic with his sister Daniela, “as we add more compute and training tasks, AI systems get predictably better at essentially every cognitive skill we are able to measure.”  The term AI scaling is used to refer to the fact that AI performance can often be improved increasing three related components:  the amount of input data, the amount of computer power (the number of processor chips and the time they are used, also called “compute”), and the number of parameters in the neural networks.  

Reporters sometimes refer to ChatGPT as an LLM, but technically that is incorrect. ChatGPT is a complex product that is built on the foundation of an LLM, but it also includes a number of other elements such as a conversation interface which makes it easy for users to ask questions.  It also includes many other sub-programs such as safety filters which prevent hate speech, sexual material and other content which violates ethical guidelines.

ChatGPT was also refined with extensive training with Reinforcement Learning from Human Feedback (RLHF).  In one example, “OpenAI paid human testers to have conversations [with ChatGPT]… and rate the quality of its replies.” In another, people were given two or more ChatGPT answers to the same question, then asked “Which is better?”  The results of these and other experiments were then used to alter ChatGPT’s responses so they would sound more like users were talking to human beings.

To become as powerful as ChatGPT, the amount of data and computing power required was staggering. In  one of the best introductory overviews of  LLMs, educator Grant Sanderson estimated that if an LLM were trained on a supercomputer network that could only complete a mere “one billion additions and multiplications every single second… it would take well over one hundred million years.”  Fortunately, today’s supercomputer networks have so much power that it didn’t take that long. 

Scaling remains a key strategy, but even Sam Altman, the CEO of OpenAI, has admitted that the “original scaling laws are working but slowing down.”  According to a New Yorker article entitled “What if AI doesn’t get much better than this?” in OpenAI’s most recent LLM release “The increase in quality [for GPT-5] was far smaller compared with the jump between GPT-3 and GPT-4.”    More generally, according to a report from the highly regarded business school HEC Paris, “AI scaling laws are showing diminishing returns, forcing AI labs to change course.”

Despite the slowdown, the scaling laws have led to a voracious appetite for training content.  Sanderson noted that if a human being tried to read all the text that was used to train GPT-3, “they would need to read non-stop, 24-7, for over 2,600 years.”   

At one level, the way LLMs work is easy to state.  As computer scientist Stephen Wolfram put it “ChatGPT is always… trying… to produce a ‘reasonable continuation’ of whatever text it’s got so far, where by ‘reasonable’ we mean ‘what one might expect someone to write after seeing what people have written on billions of webpages, etc.’”  Even more simply, Andrej Karpathy, one of the founders of OpenAI, has said “A language model is just a system that is trained to predict the next word in a sequence.” 

Come on.  Is that it?  How could that possibly enable ChatGPT to instantly produce a 20 page paper on the causes of the Peloponnesian War for a dishonest college student?

If you want to truly understand the nitty gritty details of how chatbots work, you might want to put this five minute lesson aside and start applying to computer science PhD programs.  I don’t have time for another PhD, so I just asked ChatGPT to recommend a few articles for beginners.  My favorite was called How ChatGPT Works: A Simple, No-Jargon Explanation which says, in part: “It’s not magic — it’s pattern mastery. [ChatGPT] doesn’t ‘understand’ things like humans do, but it has seen so much text that it knows:

• What sentence usually comes after another
• Which words commonly appear together
• How people ask questions
• How people answer them”

The biggest challenge to truly understanding how chatbots work at a deep level is that no one knows.  In 2026, the New York Times magazine published an article entitled We Don’t Really Know How A.I. Works which pointed out that “It has been estimated that the latest versions of Google Gemini and OpenAI’s GPT-5 contain trillions of mathematical functions…  But one cost of that improvement has been transparency. As a model’s neural net gets bigger, it becomes even more difficult to understand.” 

This black box problem is now being addressed by a number of scientists in a new AI sub-field called “interpretability.”  The Times article quoted Ellie Pavlik, one of the leaders in this field as saying, “We have made progress over the past few years, but every few months we’re deeply considering a method, and then we’re deeply considering another method.”  Finally, the article concluded “it is becoming increasingly clear that we might never have a complete accounting of why a model chooses one word… over another.”

There are also increasing challenges to the often repeated claim that, as Time magazine put it, “continuing to train AI models using ever greater amounts of computational power and data will inevitably lead to Artificial General Intelligence.”   Among those who disagree, MIT linguist and philosopher Noam Chomsky wrote in a New York Times op-ed entitled “The false promise of ChatGPT”: “It is a high-tech plagiarism system… The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response.” 

Linguist Emily Bender and her colleagues have famously written that chatbots “are for all intents and purposes stochastic parrots… stitching together sequences of linguistic forms they have observed in their vast training data, according to probabilistic information… but without any reference to meaning.”

Yann LeCunn, Meta’s chief AI scientist put it more succinctly: “Chatbots are not intelligent in any way. They are word prediction machines.”

Does this matter?  Very much to AI experts who believe chatbots will inevitably to AGI, such as the authors of the alarmist best seller “If Anyone Builds It, Everyone Dies: Why Superhuman AI Would Kill Us All.”  But the typical user is much more interested in practical questions such as the necessity of checking original sources before relying on chatbots, and the best way to phrase questions to get exactly the information you need.  If you’d like more advice on that, just ask ChatGPT.