Technology · Frontier models
How advanced AI models learn, reason and turn computation into useful work. Follow the technology from training data and model architecture to reliable applications and the economics of serving them.
Start here · The foundational guide
How frontier models work
A complete introduction to tokens, transformers, training, reasoning, mixture of experts, multimodality and inference. Then connect the model to retrieval, tools, agents and the cost of a successful task.
Go deeper · Hardware & systems
AI-chip system architecture
Follow a frontier model from calculations to accelerators, HBM, interconnects and rack-scale systems. A deeply researched guide to the bottlenecks that determine performance, with worked examples, NVIDIA–AMD–Google architecture comparisons and 31 primary references.
Go deeper · Inference & reasoning
Inference & reasoning systems
How models generate answers, use extra thinking time, manage memory and serve many users at once. Explore decoding, verification, KV caching and speculative generation, with worked examples for latency and cost per successful answer.
Go deeper · Post-training & alignment
Post-training & alignment
How instruction tuning, human preferences, reinforcement learning and verifiable rewards shape a model’s behaviour. Compare SFT, RLHF, DPO and GRPO, and explore the gap between better reward scores, useful capability and safety.
Inside the model
Model architectures explained
Understand transformers, attention, mixture of experts and long-context designs through the information they move, the parameters they activate and the memory they retain. Includes worked examples, architectural trade-offs and 24 primary research references.
From answers to actions
AI agents & tool use
How agents plan, retrieve information, manage memory and act through external tools. Understand permission boundaries, prompt injection, reliable evaluation and recovery after failures, with explanatory diagrams and 24 primary research and official documentation sources.
Four ways into the subject
What is inside the model?
Tokens become numerical representations. Attention and learned transformations use context to generate a response.
Explore transformers →How does capability develop?
Pre-training builds a base model; post-training shapes its behaviour. Data quality, feedback and compute all matter.
Follow the training process →How does a model do useful work?
Inference, retrieved information and tools turn learned capability into answers and actions in a working environment.
Explore agents and tool use →Who can earn a return?
Assess quality, latency, serving cost and the review needed to complete a task. Then connect customer value to supplier economics.
Examine model economics →The question behind the technology
A model can be technically impressive and still be costly or unreliable for a particular job. This section asks how capability is created, how it is measured and what it costs to deliver. It connects the underlying mechanisms to the decisions facing customers, developers and investors.
We use “frontier” to describe models near the leading edge of broadly useful capability when assessed. The boundary moves, and performance differs across tasks. The guide explains publicly documented techniques while identifying where commercial model details remain undisclosed.
Our analytical focus is the cost of dependable work. A lower token price is useful, but its value depends on how much computation, tool use and human review are needed to reach an accepted result.
Connect the model to the wider stack
The physical foundation is covered in AI infrastructure & data centres and semiconductors & chipmaking. Current business information comes through data platforms & analytics. Permissions and protection connect to cybersecurity.
For the commercial debate, continue to model commoditisation and open weights. Company research includes OpenAI, Anthropic, Alphabet and Meta.