Archive
Every article is kept here at its original address. For a suggested starting point, choose one of the reading guides.
Explore the reading guides13 articles, shown from newest to oldest.
Jeff Bezos's test of one-way and two-way doors maps neatly onto where to let AI act alone. The twist is that automation reshapes which door you are standing in front of, so reversibility becomes something you engineer rather than something you inherit.
Practitioners using JSON-schema enforcement, formal grammars, and regex-based decoding are doing classical symbolic AI without naming it. Naming it changes how the next round of investment gets made.
Retrieval augmented generation became the default for grounding language model outputs. For domains where being wrong is expensive, knowledge augmented generation is the next move, and the reason it works is older than the technique.
Davis, Shrobe, and Szolovits defined knowledge representation through five roles in 1993. Used as a diagnostic, those five roles will tell you where any AI agent system is likely to fail before it does.
Modern AI agents are not pure neural systems. They are hybrid systems whose reliability comes from specific symbolic investments made in the layer around the model. The pattern is selective, not universal.
AI systems do not have to erode human skill. Designed deliberately, they can preserve and even strengthen it. Six practical design patterns, and the continuum that tells you where to keep practising.
From medicine to aviation to driving, the same pattern recurs: skills that are not practised atrophy, even among highly trained professionals. The business risk follows close behind.
AI delivers real productivity gains, but relying on it to do our thinking quietly reduces the mental engagement that builds and maintains skill. Here is what the research shows, and what it does not.
Andrej Karpathy's 'LLM Wiki' pattern offers a compelling alternative to retrieval-augmented generation. Instead of re-deriving insights from scratch on every query, AI incrementally builds and maintains a persistent, structured knowledge base. The implications for organisational knowledge management are significant.
AI is widely used to process existing information. Its potential to capture expertise that was never documented in the first place is largely overlooked.
Most AI is designed to optimise tasks. But when it's built without reference to the expertise behind those tasks, it doesn't just fail to capture that knowledge; it can actively erode it.
Around 80% of AI projects fail to deliver business value. We break down the three root causes, including the tacit knowledge problem most teams miss.
Five essential questions before any AI project, including how to account for the tacit knowledge your experienced people rely on.
Earlier project pages
The earlier product and advisory pages remain available here as the site focuses on practical writing about AI and human expertise.