Artificial Intelligence
Security, governance, agents, and retrieval. The pages here are written from open standards and primary documentation - principally the OWASP GenAI Security Project, the NIST AI Risk Management Framework, and the EU AI Act - rather than summarised from coverage of them.
Topics
Hubs collect a topic area; deep dives cover one thing properly.
Follow the visual loop from a user goal to model proposals, authorised tools, observations and a verified answer. Includes a worked document-assistant example.
Read model and tool traces, reconcile uncertain writes, and distinguish a convincing success message from an observed outcome.
When to divide work, how adaptive model routing uses cascade and critique paths, and what coordination costs.
What makes a model generative rather than predictive, the two model families behind almost everything shipping today, and why the same prompt can produce a different answer twice.
What is actually established about prompting, how to spot folklore, why prompt changes need evaluation, and where prompting stops and retrieval begins.
What an eval actually is, the three families of grader and what each costs, and how far an LLM judge can be trusted with the measured numbers.
How a model asks for a function to be run, why the model never runs it, and where the boundary between a fluent request and an executed action actually sits.
What MCP standardises, the client-server split it enforces, and why that standardisation does not remove the trust problem it sits on top of.
What LangChain’s agent harness provides, why LangGraph exists as a lower-level graph runtime underneath it, and how to tell which layer a problem belongs to.
The OWASP Top 10 for LLM Applications 2026 and the AI Security Verification Standard: the risks, the twelve control chapters, and how they map onto real systems.
Why prompt injection has no reliable fix, the surfaces it arrives through, and the difference between controls that reduce attack success and controls that bound the blast radius.
The NIST AI Risk Management Framework, the EU AI Act and ISO/IEC 42001 - obligations, dates and thresholds taken from the sources themselves.
Adoption patterns, agent architectures, model risk, and what changes when an LLM is given tools and autonomy.
Retrieval-augmented generation end to end: chunking, embedding, retrieval, prompt assembly, and the production failure modes.
Context, checkpoints and long-term memory are different responsibilities. Design provenance, correction, deletion and retrieval across sessions.
Why interleaving reasoning with action outperforms planning everything up front, and what ReAct and Reflexion actually showed.
The open protocol for communication between independent agents - the Agent Card for discovery, the Task lifecycle, and where the boundary with MCP sits.
A SKILL.md file plus bundled resources - packaged procedures an agent loads through progressive disclosure instead of a growing system prompt.
The execution layer that owns the agent loop, permissions, sandbox and recovery - and how it differs from a framework and from a protocol.
Where code a model wrote runs unreviewed - the isolation primitives, the lifecycle, and how E2B, Modal and Daytona differ on what matters.
The engine that makes one replica fast (vLLM) and the control plane that decides where each request runs (llm-d) - and why prefill and decode split.
TypeSafe’s decision-only model: typed answers with calibrated probabilities instead of generated text, and where a snap judgement fits an agent loop.
Jev judgements as SQL functions: filter, rank and classify Postgres rows in plain language, with the per-row API cost and privacy trade-offs named.
Core AI terms - transformer, attention, embedding, hallucination, alignment and more - defined once and linked from everywhere they apply.
Structured data on significant AI models: capabilities, context windows, modalities, access and known limitations.
Organisations shaping the AI ecosystem: model developers, semiconductor companies, cloud providers, robotics and research.
A chronological record from Turing to the present: research milestones, model releases, regulation and industry events.
A high-level map of the AI ecosystem connecting models, infrastructure, applications, companies, research, regulation, security and economics.