10 Layers of AI
From raw data to autonomous agents — click any layer to explore
10 Autonomous agents Multi-step planning, tool use, self-correction GPT-4o · Claude · Gemini
At the frontier of AI, autonomous agents go beyond answering questions — they pursue goals across multiple steps, deciding what tools to use, when to search the web, when to write and run code, and how to recover from errors. They operate in a loop: plan → act → observe → re-plan.
Modern agents like Claude and GPT-4o can fill out forms, manage calendars, write and debug entire codebases, and coordinate with other agents in a pipeline. The key challenge is reliable reasoning under uncertainty — knowing when to ask for help rather than confidently hallucinating.
9 Reasoning & planning Chain-of-thought, symbolic logic, search trees o1 · DeepSeek-R1
Reasoning models are trained to think before they answer. Instead of generating a reply immediately, they produce an internal chain-of-thought — working through sub-problems, checking their work, and revising — before committing to an output.
This is distinct from raw language generation. Models like o1 spend more compute at inference time, exploring a search tree of possible reasoning paths. The result is dramatically better performance on math, logic, and multi-step code — tasks that require deliberate structured thought, not pattern matching.
8 Large language models Transformer architectures trained on text at scale GPT · Claude · Llama
Large language models are transformer networks with billions of parameters, trained on internet-scale text to predict the next token. What emerged — unexpectedly — was broad capability: translation, coding, reasoning, summarisation, and creative writing, all from a single objective.
The key insight is that scale unlocks capability. Models trained on 10× more data with 10× more parameters don’t just perform 10× better — they acquire entirely new abilities that smaller models lack. This is the “emergent behaviour” phenomenon that surprised even their creators.
7 Foundation models Pre-trained on massive data, fine-tuned for tasks BERT · CLIP · Stable Diffusion
A foundation model is a single large model trained once on broad data, then adapted to many downstream tasks via fine-tuning or prompting. Before this paradigm, every task needed its own model trained from scratch — expensive and brittle.
Foundation models store compressed world knowledge in their weights. CLIP learned visual concepts from 400 million image-caption pairs. BERT learned language structure from Wikipedia. The pre-training does the heavy lifting; fine-tuning shapes the final behaviour cheaply.
6 Deep learning Multi-layer neural nets: CNNs, RNNs, Transformers ImageNet · AlphaGo · DALL·E
Deep learning stacks many layers of artificial neurons, each learning progressively abstract features. A CNN trained on images learns edges in layer 1, shapes in layer 3, and faces in layer 10 — without being told what to look for. The hierarchy of features is discovered automatically.
The 2012 AlexNet moment — a deep CNN crushing the ImageNet benchmark — sparked the modern AI era. GPUs made training practical; ReLU activations made deep networks trainable; dropout prevented overfitting. The transformer architecture (2017) then replaced RNNs for sequences and enabled today’s LLMs.
5 Machine learning Algorithms that learn patterns from labelled data SVMs · Random forests · XGBoost
Machine learning is the shift from hand-coding rules to learning rules from data. Instead of writing “if pixels form a circle, it’s a zero”, you show the system 60,000 labelled digits and it figures out the rules itself. The programmer defines the model family and loss function; the algorithm finds the best parameters.
ML spans three paradigms: supervised (labelled input-output pairs), unsupervised (find structure in unlabelled data), and reinforcement learning (learn from reward signals). Classical ML methods like gradient boosted trees still outperform deep learning on many tabular datasets.
4 Statistical inference Probability, Bayes theorem, distributions Bayesian nets · HMMs · Kalman
Statistical inference is how AI systems reason under uncertainty. Rather than outputting a single answer, a probabilistic model outputs a distribution over possible answers — expressing confidence, not just prediction. Bayes’ theorem is the backbone: update your belief about the world as new evidence arrives.
Hidden Markov Models powered speech recognition for decades. Bayesian networks encode causal relationships. Kalman filters track moving objects. Even modern neural nets are implicitly Bayesian — their softmax outputs are probability distributions. Every well-calibrated AI model is doing statistics.
3 Algorithms & optimization Gradient descent, backpropagation, loss functions Adam · SGD · RMSProp
Optimization is the engine of learning. Training a neural network means finding parameter values that minimise a loss function — a measure of how wrong the model currently is. Gradient descent does this by repeatedly nudging parameters in the direction that reduces loss, guided by the gradient (the slope of the loss surface).
Backpropagation efficiently computes gradients through every layer using the chain rule of calculus. Modern optimizers like Adam add momentum and adaptive learning rates, making training faster and more stable. Without efficient optimization, even a perfectly designed architecture would never converge.
2 Mathematical foundations Linear algebra, calculus, information theory Matrices · Tensors · Entropy
All of AI runs on three branches of mathematics. Linear algebra provides the language of data: vectors represent inputs, matrices represent transformations, and tensor multiplication is how neural networks process batches of data in parallel on GPUs.
Calculus powers learning: derivatives measure how outputs change with inputs, and the chain rule links gradients through layers. Information theory (Shannon, 1948) quantifies knowledge and uncertainty — cross-entropy loss, the most common training objective, is a direct import from information theory. These three disciplines are the non-negotiable prerequisites.
1 Data & representation Raw inputs: pixels, tokens, signals, embeddings Datasets · Embeddings · Tokens
Everything in AI begins with data. A photo becomes a 3D array of pixel values. A sentence becomes a sequence of integer token IDs. A sound becomes a spectrogram of frequency intensities over time. Representation determines what patterns are learnable — a poor representation makes the learning problem harder or impossible.
Embeddings are learned representations: dense vectors where similar concepts cluster nearby in space. Word2Vec showed that “king − man + woman ≈ queen” — meaning encoded as geometry. Modern LLMs start by converting each token into a high-dimensional embedding vector. The quality of the data shapes the ceiling of every layer above it.