How AIreally works.
Here I show where my knowledge comes from: from the physics behind neural networks to the tokens that drive every AI conversation. No hype: the technology as it really works. So you can judge for yourself what is real, and our conversation can be about your business straight away.
AI began in physics.
The line from my studies to modern AI runs surprisingly straight, through systems that order themselves into a stable state.
- 01
Physics & order
My field of study is astronomy. In physics you learn about electromagnetic fields, potential distributions and systems that order themselves. At the core of many physical processes is that systems tend to move towards a stable or energetically favourable state.
- 02
The Ising model · 1920s
In the 1920s the so-called Ising model emerged in physics. This model described a network of small particles (“spins”) that influence each other and order themselves until the system settles into a stable state. Scientists later discovered that such systems resemble how learning and pattern recognition work to a surprising degree.
Just like the spins in the Ising model: the system rolls by itself to the deepest valley, the most stable, lowest-energy state. (own animation) - 03
Hopfield & energy landscapes · 1980s
In the 1980s John Hopfield used these ideas to build the first neural networks that used a kind of “energy landscape”. The network could organise itself towards stable patterns. That allowed it, for example, to store information and recognise similar situations.
- 04
Boltzmann Machines
Later, models such as Boltzmann Machines emerged, based even more strongly on physical principles such as energy, probability and stability. These ideas formed an important historical basis for machine learning and neural networks.
- 05
Large Language Models
Large Language Models as we know them today are technically far more advanced, but the underlying idea still strongly resembles these physical systems: an enormous network of connections that organises itself during training until it learns to recognise stable patterns and structures in data. That makes the architectures of modern AI and these early physical models conceptually very similar.
More powerful than ever, and energy-hungry.
Pattern recognition
Today some Large Language Models have been trained on extremely large amounts of data and can recognise almost every pattern we deal with in daily life. This pattern recognition, and the way AI works, increasingly resembles how a human brain works.
The energy problem
A growing problem in the development of AI is that energy is becoming scarcer. With every new large AI model released, energy consumption goes up enormously. For this reason, using AI such as ChatGPT and Claude is becoming ever more expensive.
AI chops language into pieces.
Using AI costs tokens. When you put a sentence into AI, it chops that sentence into pieces, and each piece is called a token.
One token costs more energy than another, but because it involves tens to hundreds of thousands, and nowadays millions, of tokens, tokens are generally a reliable way to measure your AI usage. That is also why you pay for AI usage per token. Your limits are set at a certain number of tokens as well.
Tokens become numbers, and language again.
Behind the scenes every token becomes a row of numbers: an embedding vector. The model multiplies it by its learned weights, and the result is turned back into language. Below you see that happen: from "capital france" to "paris".
But how does the model choose that answer? For every next word it computes a probability for all possible words, and (usually) picks the word with the highest probability.
Tokens in, tokens out.
For every token you put into the LLM, it starts calculating. The calculation produces tokens again, which are then turned into text. That text is the response you get from the model.
Through pattern recognition, each word in the sentence is (usually) the word with the highest probability. The tokens you put into the model are called input tokens, and logically the tokens that come out are called output tokens.
| Model | Input / 1M | Output / 1M |
|---|---|---|
| Claude Opus 4.7 | $5 | $25 |
| Claude Sonnet 4.6 | $3 | $15 |
| GPT-5.5 | $5 | $30 |
| GPT-5.4 | $2.5 | $15 |
| Gemini 2.5 Pro | $1.25 | $10 |
| Gemini 2.5 Flash | $0.3 | $2.5 |
Rates per 1 million tokens in USD · official provider prices, checked May 2026. Current rates may change.
Cheaper per token. Not necessarily cheaper per task.
The price per token has fallen sharply since 2023, and those are real provider prices. At the same time, modern reasoning and agent models use far more tokens per task.
The result: you pay less per token, but a whole task can still end up more expensive. Cost per task = price per token × tokens per task.
- Price per token
- Tokens per task
- Cost per task
Indexed to 2023 = 100, logarithmic axis. The price line follows real provider prices ($60 to about $10 per 1M output tokens); usage and cost are illustrative, with cost = price per token × token usage per task.
By minimising token usage you also reduce energy consumption, and your AI usage effectively becomes cheaper.
The workings and knowledge of these systems are the reason I find physics so fascinating.
That combination is rare: most people know either the theory or the practice. I deliberately move between both.
On one side I study the physical and mathematical models, and on the other I work on their direct application to business cases.
My goal is to bring this knowledge to businesses, to gain insight into their processes together, so that we can put these kinds of systems to work for them.
For you that means: I understand not only what an AI system does, but also why it works, and therefore where it does and does not fit within your processes.
Would you like to apply this to your own processes?
Email me. One paragraph is enough. I read, think it over, and answer honestly, even if I think AI is not the right solution.