AI Is Not Science. It’s Alchemy. Here’s the Evidence

Isaac Newton was a brilliant scientist who discovered the law of universal gravitation. He also formulated the laws of motion and invented calculus. Without these achievements, modern physics, engineering, and space exploration would be unimaginable.

But few people know that he dedicated a significant portion of his scientific pursuits to alchemy.

Was he successful in that field?

No, because alchemy turned out to be a pseudoscience.

Newton and other exceptionally bright and talented people didn't know this — and couldn't have known.

They were pioneers. They were walking in the dark, trying to find their way through the darkness to distinguish truth from bias.

Centuries later, we benefit from their experience and honor them. We owe them our gratitude, in part, because they have already separated the wheat from the chaff for us.

But in the 21st century, alchemy has returned to our lives. This time, it has taken the form of mainstream AI.

Does history teach us nothing?

What the LLM Adepts Believe

Our "enlightened" age is remarkably similar to Newton's time in one regard: we once again want to believe in miracles. For three years, we've been told that AI is about to bring about superintelligence that will transform the face of humanity.

But this is all mere speculation, aggressively pushed by the CEOs of AI labs and hyperscale companies, various "experts," and the media. These claims are akin to the alchemists' promises to create the "Philosopher's Stone." It was impossible because alchemists were guided by speculative premises, not scientific theory.

Look at how AI labs justify their promises, and you'll see that history is repeating itself.

The uncomfortable truth is that today's AI is not based on scientific knowledge. It is based on a cult that asserts the power of neural networks driven by brute force.

Neural networks have been around for a long time, and at first, they didn't seem like a silver bullet. They failed to produce a breakthrough in AI and couldn't prevent the field's second "winter" in the late 1980s and early 1990s.

They only proved effective once it was possible to train them on massive datasets using enormous computing power. Thus, their success in implementing AI is inextricably linked to scaling.

The Bitter Lesson of Brute Force

Behind the idea of scaling lies an entire philosophy, codified in Richard Sutton's famous 2019 essay, The Bitter Lesson.

Sutton is an influential AI theorist and one of the founders of reinforcement learning. He is a 2024 Turing Award winner and co-author of the classic textbook Reinforcement Learning: An Introduction (2018).

The Bitter Lesson argues for abandoning any approach to creating AI other than the brute force of computation.

The title is telling. The lesson is "bitter" because Sutton, as a scientist, understood that any concept should ideally be based on human knowledge. Yet, he believed that human efforts to bake knowledge into AI were doomed to fail because, in the long run, they lose out to brute force. He writes:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

Sutton did not provide a scientific basis for this claim. He simply presented it as a given and declared that all alternative attempts were a waste of time. Essentially, his message declares all other approaches a priori invalid.

Notably, The Bitter Lesson was written several years before the advent of ChatGPT. However, once LLMs went mainstream, Sutton himself became a critic of the approach.

In 2025, he officially expressed the opinion that LLMs are a dead end. In his view, they merely mimic thinking rather than figuring out what is actually happening in reality.

Yet, the followers of LLMs were not bothered by Sutton's criticism. It doesn't negate his assertion regarding the indispensability of brute force; thus, the philosophy of scaling continues to justify using deep neural networks to build AI.

Neural Networks

The next component of AI alchemy is the Neural Network.

Historically, their developers have pointed to the way the brain works. This is a point frequently made by the "Godfather of AI," Geoffrey Hinton.

But a problem arises: the "brute force" dogma rejects human-crafted knowledge, including knowledge of how the brain works.

As a result, these two pillars of the modern AI paradigm are in conflict. One legitimizes neural networks through human knowledge; the other denies that knowledge any credibility.

This "epistemic double-entry bookkeeping" undermines the integrity of the entire AI structure, from Deep Learning to LLMs:

But the problems with the mainstream approach don't end there. The claims of neural networks themselves do not hold up to scrutiny.

First, when Hinton defends their biological plausibility, he relies on a classical theory of how the cerebral cortex works — one that is now outdated. Its basic idea can be illustrated like this:

Here, the brain is represented as a multi-layered network of neurons. Each subsequent layer extracts increasingly complex features from incoming information: first lines and angles, then individual shapes that make up the features of an object, then the features themselves, and finally, the object.

The problem with this theory is that it provides no way to understand the morphology and composition of objects. Consequently, modern systems try to classify them simply by a set of features.

This is a fragile and unreliable method. A different camera angle, a change in lighting, or an obstruction by another object, and the recognition system fails.

• Alchemists operated on a similar principle. They believed God left external "signatures" on every object, which could reveal its hidden nature and purpose: — A walnut looks like a brain → it cures brain diseases. — Liverwort looks like a liver → it treats the liver. — A plant is red → it helps with blood disorders.

The LLM adepts have, of course, gone further than their alchemist predecessors. They handed the task of recognizing these "signatures" to machines. But they similarly failed to find a way to determine the internal properties of what is being recognized. A neural network merely maps observed features against the statistical combinations it saw during training.

The next problem lies in determining the validity of the output.

How does the system know it has correctly identified an object — that what it recognized as a cat or a dog is actually that animal and not something else?

The short and honest answer: it doesn't.

It doesn't know there is actually a cat or a dog in front of it. It only calculates that the observed set of features most likely corresponds to one of the classes present in its training data.

This is not knowledge in any meaningful sense.

Backpropagation

Then, the "crown jewel" of deep neural networks is backpropagation, a point of pride for Hinton and many of his colleagues.

This algorithm reduces the error rate on training examples by automatically adjusting billions of model parameters. Thanks to backpropagation, these models demonstrate impressive results on massive datasets.

However, behind this elegant math lies the same principle seen in the previous cases. Backpropagation does not explain to the system why its output was wrong, and it does not teach it to understand the object.

It merely adjusts the weights numerically so that the statistical error is smaller the next time.

• Of course, the alchemists had a similar approach. They compiled (endlessly) "corrected recipes." If the result was unsatisfactory, one should: — Heat it more; — Add more sulfur; — Increase the heating time; — Repeat seven times (and if it fails again — repeat forty times).

Sometimes, it actually worked. If they had had modern equipment, they likely would have had more success.

Nevertheless, they would still lack an explanation of why the recipe worked. Current AI adepts lack an explanation of how an endless sorting of pixels or tokens relates to the internal essence of an object — and what any of this has to do with intelligence.

This is exactly why training neural networks is impossible without humans. They are shown millions of pre-labeled images where it is already known what is a cat, a dog, a car, or a tree.

After every response, an external algorithm compares the network's output to the correct label and calculates the error. Only then does backpropagation adjust the parameters so the probability of a similar error is lower next time.

At no stage does the system answer the question of why its output was wrong. It doesn't analyze the cause of the error or build a causal model of the object.

It simply reproduces statistical dependencies between input data and the correct answers provided by humans with increasing precision. Essentially, it utilizes human knowledge rather than learning causality.

Finally, a mechanism similar to backpropagation has never been found in the brain. Thus, the claim that the neural network approach is consistent falls apart before our eyes: on one hand, they claim inspiration from the brain; on the other, their key method has nothing to do with how the brain functions.

Conclusion

The modern approach to building AI is not a coherent concept. It is a combination of methods that lack a shared theoretical foundation.

Sutton's belief in the superiority of brute force remains unproven. He simply generalized observations spanning several decades, but he never explained why these observations are relevant to all AI research fields or why they will remain so forever.

Neural networks are not based on the way the brain works. Their proponents use neurological rhetoric rather than evidence. Even the superficial similarities one might find (with a stretch) refer to an outdated scientific paradigm.

The key component of neural network efficiency — backpropagation — has no analog in the brain, a fact that even Hinton was eventually forced to admit.

In the end, mainstream AI resembles alchemy precisely because it lacks a unified theory. Such a theory should be a theory of intelligence, but none exists here. Instead, it stitches together disparate ideas whose origins and foundations are fundamentally different.

Here, we see the unproven faith in the universality of brute force sitting alongside rhetorical references to biology, statistical optimization methods, and empirical engineering recipes.

The fact that individual elements "work" is taken as confirmation of the entire construct. However, there is no scientific justification for why this specific combination should lead to the realization of intelligence.

This is the state of affairs in the field that, according to AI adepts, will change the entire world in the coming years.

So, we can all relax and take our time preparing for the arrival of the Philosopher's Stone.

After all, if they promise it to us, how could we not trust them?

Enjoyed the article? Subscribe to my updates to stay in the loop on what's happening behind the scenes of the AI industry.

You may also like these stories:

• A Terminal Diagnosis for AI: Here's Why They're Desperately Trying to Hide It
• Not Quite Science War Around AI: The Real Reason Why Geoffrey Hinton Hates Noam Chomsky
• Simple Proof That LLM Is a Stochastic Parrot That Doesn't Understand the Meaning of Text

You can also read my Selected Articles on Artificial Intelligence for Thoughtful Readers.