Recently, a client asked me the following: If I had to explain the AI industry in one go, what would you give me? This had to be explained in a high-level, very intuitive way for anyone who’s not quite as deep into AI as I am to follow.

In particular, answer the following:

In this piece, we’re doing just that. It’s my longest article ever, but I’ve genuinely never packed more information and insights into a single piece. Ever. It was a hustle (I honestly don’t know when I decided it was a good idea to put so much effort into something that can be purchased for $20/month, but I guess I really dislike doing things halfway).

Alas, I hope you enjoy it.

Let’s dive in.

Understanding the technology is not just about memorizing the key terms; it’s about understanding why it looks the way it does, why it needs the hardware it uses, and why it takes the particular product shape it does.

An AI algorithm is just a method of processing data. By ‘processing’ we mean capturing the underlying patterns in data to make predictions about it.

The constant definition you see in the wild is that it learns patterns from known data to make predictions on unseen but similar data.

Say you want an AI to predict housing prices, trained on millions or billions of housing data points with known prices. Eventually, it starts picking up regularities in the data, such as “houses with many rooms tend to be pricier,” “postal codes in this area tend to be cheaper,” or “marble countertops are often present in expensive homes.”

With these captured patterns, the AI model can make predictions about new homes it has not seen before. It doesn’t know the price, but if they see that the house has 7 rooms, a postal code from a rich suburb, and marble countertops, the model can infer that it should be priced high.

Those correlations are what AIs learn.

Traditionally, however, in the era of Machine Learning (pre-2010s), we lacked the necessary compute (physical hardware) to feed models the humongous amounts of data we feed to modern models today, which allows AIs to find all these patterns.

Therefore, we performed what’s known as ‘feature engineering’: we would run statistical analyses on the data, identify which variables mattered, and provide the model with the answers in advance, making what the AI learned highly relevant to what the engineers wanted.

Using our housing analogy again, we didn’t have trillions of homes to share data with the AI so that it can autonomously figure out “what mattered”, so we would previously run statistical analyses like regressions or correlations to identify the key variables (e.g., room count, postal code, previous prices, and whatever most likely determined the price of the home) and basically build that model around those assumptions.

But in the 2010s, the Deep Learning revolution changed AI forever, taking us to where we are today.

Although the principles of deep learning are older than basically every single reader of this newsletter or close, over the last 15 years or so, the world started to see more compute become available, mostly via GPUs (Graphical Processing Units), allowing researchers to tap into more and more data to feed to the AI models, which, as you may have guessed by now, is what all this is about: feeding data to AIs that capture the patterns in that data and can make inferences about them; as much data as one could get their hands on.

This led to the ‘Big Data’ era, a term you've likely heard before, which opened the door to a particular type of AI algorithm called neural networks, also known as Deep Learning.

But why? The short answer is the Universal Approximation Theorem, the idea that neural networks can approximate any continuous mathematical function to arbitrary accuracy.

In layman’s terms, if a relationship between two variables can be represented by a continuous function, neural networks can learn it.

But what do we mean by mathematically represented? By that, I mean the AI can perform mathematical computations to predict a possible answer, and that answer can be mathematically measured to tell us how good or bad it was, and thus used as a learning signal.

LLMs are a great example of this, but it applies to every single AI model on planet Earth. The LLM sees the sequence “What year was Einstein born?” The ground truth answer is 1879, which the LLM does not know.

The LLM then performs a series of computations, resulting in a ranking of possible answers by likelihood, assigning the year 1879 a probability of 30%. Of course, the correct answer would have been to give that year a probability of 100%, so we have a mathematical representation of the mistake: the model was off by 70%.

That is what we mean by mathematically represented: you can measure how good or bad a prediction was using maths. And if that loss is measurable, you can train the AI to minimize that loss; it’s mathematically guaranteed.

And there you go, you know understand what ‘AI training’ actually means: the AIs make predictions, those predictions are measured by how good or bad they were, giving us a loss, and that loss is used as a learning signal (i.e., the AI knows what direction it needs to go in order to reduce the loss).

In our case, the next time, the model will assign a higher probability to ‘1879’ and, across trillions upon trillions of predictions, the LLM learns to assign high probability to the correct words.

Importantly, the ability to measure how good or bad a prediction was is called ‘verifiableness’. This is important because if you understand this concept, you immediately know where AIs are expected to be good and where they are not.

For example, look at these two sentences:

Both could be responses coming from an LLM. But how verifiable are they? The former is clear; the LLM output was accurate because the answer to 5 plus 6 is indeed 11. That is a sign that the LLM is improving in maths. If the model had answered 10, we also know the model is wrong. That is a sign that the LLM needs more work.

But the second one is different. If you’re trying to make the LLM better at writing or storytelling, how good or bad is the latter sequence toward that goal?

Yes, we see good grammar, which is a factor in good writing. What makes writing great? At best, that’s subjective. That is not a good example of a verifiable data point; we really don’t know, mathematically speaking, how good that output was.

This is the verifiability problem, and it’s crucial because it’s a hard constraint for AIs: they can only learn what is verifiable.

If you understand this, you know how to protect your future because, to me, verifiability is the only obvious answer to the question: “How do I protect my job from AI?”

I can’t promise you that mathematicians and software developers, both in highly verifiable domains, won’t see meaningful changes in their work over the next few years, but I am very comfortable telling you that an artist’s job won’t change much.

Of course, AI does impact artists in meaningful ways, especially on the lower bound of work; you’re no longer as capable of selling some type of art that AIs can more or less copy, but high-end work, unique and original, is not something AIs can do anytime soon because they don’t have a “mathematical hill to climb.”

Other examples include sales: What makes a good saleswoman? You can sense when someone is great at sales, but please try to represent their sales skills using maths. Good luck.

AIs do have a way to reasonably address some of the issues with the non-verifiable nature of certain skills, which leads us to the two types of learning mechanisms.

In AI, there are two ways you train models: imitation and reinforcement.

Imitation refers to the technique of exposing an AI to an absurd amount of data and having it replicate it to the letter. But by doing this and adding inductive biases that induce compression (I won’t get into this), models learn to imitate us.

This is how ChatGPT learns not only the language but also how to speak back to you. It does so by imitating humanity’s digital corpus to the dot. This is how AIs learn to imitate Shakespeare, or cite Martin Luther King’s Lincoln Memorial speech, “I Have a Dream,” by memory.

However, as it has been trained over the entire Internet, the outcome of this is like a mediocre “human,” the average of humanity’s digital content, an AI model that becomes the embodiment of the average word on the Internet.

But current top models include an additional step called reinforcement that changes things. Here, instead of giving them a lot of data to copy, we now give them goals and tasks to solve and let them try… a lot.

And when these AIs “stumble” upon a solution we can verify as correct, we reinforce that behavior so it’s more likely the AI will repeat it.

Once the model reaches the solution, the entire “reasoning” that got it there gets reinforced, making it more likely the model repeats it.

This introduces its own problems, but it really pushes AIs to reach new heights in those areas, as AIs learn reasoning patterns that apply well to maths or coding structures, leading to successful outcomes.

The reason this is so effective is simple: they become problem-solvers at scale, and when you can verify whether what they are doing is correct, you can verify progress and thus make progress.

Therefore, this technique, known as Reinforcement Learning (RL), is what has turned AIs into tools that are actually useful and even discover new maths… but only in those areas that can be verified.

Therefore, most humans using AIs to generate content are behaving as if they had hired a really inexpensive but very bad copywriter to handle their digital blueprint.

For all these reasons, if you use AI to write, your content was, is, and will be mediocre for the foreseeable future because OpenAI and others genuinely don’t know yet how to train great AI writers, screenwriters, poets, film producers… and with absolutely zero evidence that we might be nearing a way to solve that problem.

Ok. So, as of right now, the summary is that AI captures patterns, and that neural networks, thanks to the emergence of compute at scale, which lets us feed AIs trillions of data points, have become the main choice for building AI systems.

But which of these neural nets rules? You probably heard the name: the Transformer.

During the Renaissance, there was a specific mathematical concept known as the Golden Ratio, which many artists of the era believed would make their work aesthetically pleasing when proportioned to it. Centuries later, you would see painters like Salvador Dalí still doing the same.

In AI, there’s a similar admiration for a concept known as attention, which sits at the heart of modern AI architectures, notably the Transformer.

Additionally, in AI, we also have a thing called the ‘bitter lesson’. First described by Rich Sutton, it is the “bitter” realization that, at the end of the day, the best humans can do in our AI aspirations is to… get out of the way.

That is, it’s not about finding the most clever, complex heuristic we can find to train models. Instead, the “best model architecture” is the one that allows the model to “see” more data and thus requires more compute, which usually translates to very simple architectures that scale really well.

And the Transformer, the architecture we have today underneath most frontier AIs, is the perfect example of this.

The architecture that underpins products like ChatGPT is stupidly, almost insultingly, simple.

At its core, a Transformer is just a concatenation of Transformer ‘blocks’ that perform several linear transformations to shape the model’s internal “belief” about which word comes next.

But instead of indulging in esoteric descriptions like this one, which we can both pretend to understand but in fact don’t, I always like to explain these models more intuitively.

And once you see this, it’s like you’ve magically understood AI algorithms.