When we say “AI”, we usually mean not one thing but a whole stack of layers. At the bottom are chips and electricity, in the middle the model and its training, at the top applications, tools and the people who decide what it is all used for. Understanding these layers helps answer questions like “why does the chatbot give great advice one time and poor advice the next?” or “what exactly is new about a new model?”
From the bottom up
Hardware. Training large models runs on thousands of graphics chips in data centres and uses a lot of electricity. That is why computing power is one of the main limits — and costs — of the whole field.
Data and training. The model learns from vast amounts of text, images and code. The basic task is surprisingly simple: predict which piece of text (token) comes next. Details: How AI works.
Model. Training produces a neural network with billions of parameters — “knobs” in which learned relationships are spread out. The model itself is just a set of numbers; it does nothing until someone runs it.
Fine-tuning. A raw model can continue text but cannot be a useful assistant. Fine-tuning teaches it to answer questions, follow instructions, decline dangerous requests and admit uncertainty — using examples, human ratings and tasks whose results can be checked automatically.
Reasoning. Newer models have also learned to “think” before answering — break a task down, try an approach, check it and start again if needed. We look at this layer in detail below.
Application. What you talk to in a browser or on your phone. Around the model it adds an interface, conversation history, possibly memory across conversations, and safety filters. Two products using the same model can therefore behave quite differently.
Tools. A model can be given a search engine, a calculator, access to documents or an environment for running code. It then does not have to “know everything by heart” — it looks information up or computes the result. The quality of the answer still depends on how well it uses the tool.
Agent. When a system not only answers but plans and carries out a sequence of steps towards a goal, we call it an agent. What matters most are its permissions: read only, or also write, send, pay? More: AI agents.
People and rules. At the top are those who set tasks, check results and carry responsibility — and laws such as the EU AI Act. This layer decides what the abilities of the floors below are actually used for.
Why it matters
When AI makes a mistake, it helps to ask which layer it came from. Had the model never seen the thing? Did the app fail to pass it an important document? Did search find a bad source? Did the agent have unnecessarily broad permissions? Each layer is fixed differently.
Reasoning: when a model thinks first
The fast answer and the slow one
Try a puzzle: “A bat and a ball cost £1.10 in total. The bat costs £1.00 more than the ball. How much does the ball cost?”
Most people first blurt out 10p. But then the bat would cost £1.10 and together they would cost £1.20. The right answer is 5p (bat £1.05). The psychologist and Nobel laureate in economics Daniel Kahneman, in Thinking, Fast and Slow (2011), described this as the difference between fast, intuitive thinking and slow, careful thinking.
A classic language model answers “fast”: token by token, without the chance to stop and go back. Reasoning is the attempt to give it that slow thinking.
Step by step
As early as 2022 researchers showed that if you ask a model to work step by step and write out intermediate results, it solves word and logic problems much better. The method was named chain of thought. (Wei et al., 2022)
Why does it work? A model “thinks” only by producing text. Every intermediate step it writes becomes part of the context from which it computes the next token — so it can “jot down” intermediate results instead of having to do everything in a single step.
Models that learned to think
The next step came when models began learning to think by themselves. In September 2024 OpenAI introduced its o1 model (Learning to reason with LLMs); in January 2025 DeepSeek published a detailed description of its R1 model (DeepSeek-R1). The principle is similar in both:
- The model gets a task whose result can be checked — a maths problem with a known answer, a programming task with tests.
- It first “thinks”: writes out its approach, tries paths, notices mistakes.
- For a correct result it gets a reward; this kind of training is called reinforcement learning.
- After many repetitions it learns the approaches that lead to good results: break the problem down, check intermediate results, go back when something doesn't add up.
Interestingly, nobody told the model how to think. Longer and more careful approaches emerged because they led to the reward.
More thinking = more computation
Reasoning brings a new dial: how much computation to give the model when answering. For hard tasks, longer thinking often leads to better results — sometimes giving an improvement comparable to a larger model. (Snell et al., 2024) It has a price, though: the answer takes longer and costs more energy and money.
Where reasoning helps — and where it doesn't
- Helps: maths, programming, logic problems, planning, analysing a complex document — anywhere you can proceed in steps and check intermediate results.
- Doesn't help: when the model doesn't know a fact. Thinking cannot replace missing information; a long argument may just wrap an error more convincingly. For facts, tools like search are better.
- Unnecessary: for simple questions it just slows things down.
Beware of the “thoughts” on screen
Some apps show what the model is “thinking”. It looks like a view into its head — but it need not be a faithful record of what actually produced the answer. Research has shown that models can write a convincing justification unrelated to their actual decision. (Turpin et al., 2023) A written-out approach is useful for checking, not as proof.
What this means for superintelligence
The layers show that progress need not come only from a bigger model. The capabilities of the whole system also grow through better thinking, tools, agents and connections to other systems. Equally, safeguards can be built into every layer: limited permissions, checking results, human approval of important steps. More: AI safety.