Justifying an oversimplification

I want to start by saying the LLM technology and tools that we have in 2026 are truly mind-blowing. The experience of using Anthropic’s Fable for the 5 minutes that it was made public in June 2026 seemed like a huge jump in progress from even a recent LLM like Opus 4.6.

A term that I’ve heard a lot (and I like to repeat it) is that the thing that is called AI in 2026 is just a really fancy auto-correct. That’s a huge, maybe flagrant, oversimplification. But the way I see it is that if people can call something “Artificial Intelligence” when it’s actually not, then I’m allowed to use a flagrant oversimplification of really fancy auto-correct. The AI models (LLMs) that are popular today are powerful pattern-completion engines. Some researchers argue that if the statistical patterns are complex enough, understanding may emerge from them. This power of matching and completing patterns looks a lot like understanding. But the problem is that this isn’t the same as an understanding of reality.

There are three kinds of lies: lies, damned lies, and statistics.

Training

The training of an LLM is to predict the next token. There is a parameter called “temperature” that will control how deterministic/random the output is. The lower the temperature (like 0.1), the more repetitive and boring it is. The top token will nearly always be selected, which leads to predictable results. The higher the temperature (like 1.5) means that responses can be more creative. But it also means they can be more chaotic. This “predict the next token” feature is also why LLMs have the “fake it ‘til you make it” attitude and have total confidence in things they have made up. We call it hallucination and see it as a bug. But to an LLM it is 100% accurate - it is the next token to them, so of course it is 100% accurate. It has no real-world experience, no way to know what “good” looks like unless we tell it what “good” looks like.

You may have seen the word “stochastic” used about LLMs. Stochastic means that something that is governed by probability. It involves randomness and chance. I mentioned the temperature earlier. This is the level of how stochastic the model is. A temperature setting is mandatory. If you set the temperature to 0, that means the model is fully deterministic. It’s still doing probability calculations, but now it’s always going to select the peak distribution rather than sampling from it. You won’t get interesting responses. A use case for setting the temperature to 0 would be where you are working with structured data and extracting data from a document. There’s only one right answer in these situations. You don’t want any randomness. Another use case would be in code generation where the code you generate is simple boilerplate code. Increasing the temperature will also increase the variation in code and that can introduce a bug.

This is the core of my argument of why AI is just really fancy auto-correct. For example, you type in “hellk” in your phone, and it will probably guess you meant “hello” and then suggest a next word to use. On my phone it suggested “Kitty”. I assume it thought I meant to type “Hello Kitty” which is something I don’t think I have ever typed. What has happened is that some reasoning of statistics happened and came to the conclusion that “hellk” is a lot like “Hello Kitty”.

Not just statistics

However, there is more to the argument than just probability calculation. There is a deeper, almost philosophical argument.

Yann LeCun is commonly referred to as one of the “Godfathers of AI”. He’s one of the architects of modern deep-learning. He’s a professor that won a Turing Award, is the creator of the OCR system (Optical Character Recognition) used in computer vision and a former Chief AI Scientist at Meta. He’s stated publicly that the current state of LLMs cannot achieve Human Level intelligence. That’s a big claim. He argues that LLMs lack a persistent world model. They cannot plan how to meet a goal. They cannot reason about physical reality.

To go further on this philosophical theme, there is another Turing Award winning scientist called Judea Pearl. He has written about how the understanding of causation is important. Pattern recognition is the 1st of a 3 stages. LLMs today are at the 1st stage where they can understand probabilities. This is also called the causal explanations stage and it makes up most of the knowledge we possess as sentient beings. Stages 2 and 3 need “causal understanding”. This allows us to control and change our environment. This involves seeing, learning from experience and the ability to imagine. As sentient-beings, we can combine these multiple variables into “what would happen if…” scenarios. However, because an LLM is only at the 1st stage, it cannot do stages 2 and 3. An LLM will learn that certain words follow other words, but they have no internal model of why.

Why pattern matching works

LLM-assisted coding harnesses are an interesting example of this. They are fantastic tools, and many seem like they can handle a large amount of coding tasks. For example, I could be working on a project using a cloud provider and the coding tool will suggest adding a permission that I need. This seems like it has read my code, understood what I am doing and thought that it needs to create a permission. But it hasn’t really. It just has done enough pattern matching to know that when someone is creating an API gateway in a cloud provider, then it’s statistically likely they need to add a permission and a policy in their environment. Coding is usually well-suited to pattern matching. A lot of the tasks are structured, learnable and variations on existing patterns (ie. a low temperature). But as soon as you go outside of the common patterns (ie. increase the temperature), the code starts to become more chaotic.

A Thought Experiment

Another way to understand this is to think of a person in a room. They have no understanding of Chinese but all they do is receive letters in Chinese characters. They follow a rule book to create responses and then give the notes back. To anyone outside of the room, it looks as if the person inside the room understands Chinese. This thought experiment, called The Chinese Room, is from 1980 but is still valid today when describing how an LLM works.

All of these concepts are important to understand when using AI coding tools. If you know that an LLM will pattern match rather than reason, then it changes how you use it. You’ll know not to put too much trust in the output of the LLM. You know that you need to verify edge cases especially in AI generated code. This is how bugs can slip in.

You hear “Air travel’s too slow”. New York to California in 6 hours. That used to take 30 years! - Louis CK

The quote above is from Louis CK. People complain about modern technology but it’s actually amazing. LLMs suffer this fate often too I think. A lot of it is because they’re over-hyped. This technology has opened up opportunities for people with access to an LLM and vastly increased productivity. I don’t want to get too deep into the downsides like the significant environmental impact of training and running models. Or the “arms race” that is happening for a country to claim that they are the winners of “AI”. Anyway, I attribute a large part of that to human nature - greed and hunger for more, whatever the cost. But in the context of the history of human beings, 2026 isn’t the worst time to be alive.