AI Systems
Models for building with machine learning and large language models: where learned behaviour beats code, how to keep agents simple, safe and measurable, how to plan for models that keep improving, and why AI pays off only when the work is redesigned around it. Includes ideas from people at OpenAI and Anthropic.
- No. 099
The Bitter Lesson
General methods that scale with computation beat clever, hand-built knowledge — eventually, and by a large margin.
- No. 100
Software 2.0
Some software is no longer written as code but learned from data; the dataset and the objective become the program.
- No. 101
Scaling Laws
Model capability improves predictably as compute, data and model size grow, so plan for the models of next year, not today.
- No. 102
Workflows Before Agents
Use the simplest LLM design that works; add agent autonomy only when a problem truly needs it.
- No. 103
Context Engineering
A model's context window is a limited budget; curate the smallest set of high-signal tokens it needs for the task.
- No. 104
Context vs Behaviour
Diagnose an LLM failure first: missing knowledge needs better context; wrong behaviour needs better prompting or tuning.
- No. 105
The Jagged Frontier
AI is excellent at some tasks and unreliable at others of similar difficulty — and the boundary is invisible until you test it.
- No. 106
The Lethal Trifecta
An AI agent that combines private data, untrusted content and a way to send data out can be tricked into leaking it.
- No. 107
The Autonomy Slider
Give AI as much autonomy as you can verify quickly — and let people move the slider up as trust grows.
- No. 108
The Dynamo Lesson (Productivity Paradox)
A general-purpose technology pays off only when work and organisation are redesigned around it, and that takes years; dropping it into the old layout gives little.
- No. 109
Evals as Fitness Functions
AI behaviour is probabilistic and shifts with every model, prompt and data change, so check quality with automated evaluations on real cases, run on every change and tracked over time.
- No. 110
Machine-Usable Interfaces
AI agents can only automate what they can reach, so expose business capabilities as documented, permissioned, stable interfaces rather than screens built for people.
- No. 111
Moravec's Paradox
What is hard for people is often easy for machines and what is effortless for people is often hard, so intuition about which tasks AI will find easy is unreliable.