The Architecture of Illusion: Breaking Down Models, Transformers, and Agents

Modern AI has to be understood without anthropomorphic metaphors. We are not talking about “thinking” machines in the cognitive, biological sense, but high dimensional statistical optimisation. The current paradigm is characterised by a triad:

1. The Model (the static manifold)
2. Transformer (the mechanism of feature extraction by means of attention)
3. Agent (the bridge to utility through iterative feedback).

The Model: The Geometry of Static Prediction

Imagine an LLM as an unimaginably detailed cartographic map of all human expression. Simple road maps are like traditional machine learning models: they only show the major highways for specific, well-defined tasks like detecting spam or predicting house prices. An LLM, instead, is a topographic survey of every square inch of a digital landscape. It doesn’t “know” what a mountain is, it just knows that if you are at point A (the word “The”), the odds of going to point B (the word “quick”) are much higher than going to point C (the word “elephant”).

It doesn’t “know” what a mountain is; it just knows that if you’re at point A (the word “The”) the probability of going to point B (the word “quick”) is much higher than it is to point C (the word “elephant”).

The “Hot or Cold” Method (The Math of Training)

When we talk about a model “learning”, we don’t mean a classroom experience. We’re talking a big machine game of “Hot or Cold.”

Picture a blindfolded archer aiming for the bullseye of a target with billions of rings. The “bullseye” is perfect, factually correct, human-like language. The archer begins by shooting arrows completely at random. When they miss, a referee says, “You are this close to the target. Then the archer alters their aim by the smallest fraction of a millimetre.

We do this billions of times, across the whole library of human text. The archer, the model, finally hits the mark. They don’t learn how to “be” an archer, they just learn the exact physical motions required to hit that target based on the statistical average of all the arrows they’ve ever shot. This is what we call “optimisation”. We are simply reducing the margin of error smaller and smaller, so that the model’s guesses are statistically indistinguishable from human writing.

LLMs and Traditional ML Models

Think of these specialists as traditional ML models. They are made for one thing, like a calculator is made for arithmetic. To make them work, you need to feed them “features” — specific, human-identified labels (“is this email short?”, “is this price high?”)

LLMs: They are generalists. They do not need us to tell them about the features. They find them out themselves. The model sees enough data to learn that “ grammar “ , “ logic “ and “ facts “ are patterns that go together . It automates the search for those patterns, making it much more flexible, but also less predictable.

Models LLM/ML as probabality models

II. The Transformer: Attention

Before the Transformer, AI read text like we read a scroll, left to right, one word at a time. If the sentence was too long, the AI would “forget” the beginning by the time it reached the end.

This changed with the Transformer, which gave the AI a “highlighting” capability. When it sees a word — it doesn’t look at the word alone. It highlights every other word in the sentence at the same time, picking out which ones are relevant to the current one.

Imagine you’re at a noisy cocktail party. You are addressing one person but you are subconsciously keeping an ear on other voices in the room. If someone in a corner calls your name your brain will register that immediately.

That’s “Attention.” As it reads a word, the model scores each of the other words in the sequence for how “relevant” they are. If the word is “ bank “ the model looks at the context — words like “ river “ or “ money “ — to decide which “ bank “ to process . It does this for every word, all at once, building a map of meaning that is richer and more connected than old, linear models could ever hope to deliver.

What is meant by attention

III.The Agent: The Bridge

A raw LLM is a completion engine. It completes your sentence. It’s not an agent. To turn it into an agent, we need to put the model in a feedback loop. Think of the LLM as a very smart engine, but the “Agent” is the steering wheel and driver.

The Loop of Feedback

An agent is simply a structured loop:

Observation: The agent observes the current state (e.g. “The user asked for a flight to Tokyo”).

Reasoning: The engine (the LLM) predicts what the next step should be (e.g. “I should search for flights”).

Action: The agent takes that step, e.g. invoking a search tool.

Feedback: The agent sees the result of that action and continues the loop.

The Agent appears intelligent because it appears to have an objective. But remember, the agent does not want to book a flight. The agent doesn’t want Tokyo. It’s just following a set of programmed instructions that says, ‘If you see a query, generate a step, execute it and see if it works.’

In the case where the odds are in favour of the next right move, it works. If it does not work it is a statistical error in the chain.

We are not building sentience; we are building increasingly complex mirrors of our own collective output. The Transformer’s ability to map structural dependencies is mathematically superior to anything we have built before, but it remains a stationary statistical process. The agent is a control loop attempting to constrain the probabilistic output of the model into something useful.

Until we move beyond simply predicting the next most likely piece of data, we must treat these systems as probabilistic engines, not reliable peers. It will save you significant time in debugging your expectations.

References

[1] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. Cambridge, MA, USA: MIT Press, 2016.

[2] C. M. Bishop, Pattern Recognition and Machine Learning. New York, NY, USA: Springer, 2006.

[3] N. Tishby and N. Zaslavsky, “Deep learning and the information bottleneck principle,” in IEEE Information Theory Workshop (ITW), 2015.

[4] A. Vaswani et al., “Attention Is All You Need,” in Advances in Neural Information Processing Systems (NeurIPS), 2017.

[5] K. P. Murphy, Probabilistic Machine Learning: An Introduction. Cambridge, MA, USA: MIT Press, 2022.

[6] D. Jurafsky and J. H. Martin, Speech and Language Processing, 3rd ed. draft. 2023. [Online]. Available: https://web.stanford.edu/~jurafsky/slp3/

[7] S. Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models,” in International Conference on Learning Representations (ICLR), 2023.

[8] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Nature, vol. 323, pp. 533–536, 1986.