Essays on applied AI, engineering and organizations

What does a model actually receive, and what follows from how text is cut into tokens?

Reviewed · version 3 · 11 min read

Why does each lab's model behave differently, and why does last year's prompting advice stop working?

Reviewed · version 3 · 11 min read

Where do convincing errors come from, and why do different fixes solve different problems?

Reviewed · version 3 · 11 min read

Why did coding agents improve faster than almost anything else, and which kinds of work come next?

Reviewed · version 3 · 9 min read

How do a training objective and enormous data produce abilities nobody programmed one by one, and what remains open about 'understanding'?

Reviewed · version 3 · 11 min read

What do reasoning models do differently, and when is extra thinking worth its cost?

Reviewed · version 3 · 11 min read

If writing code costs a fraction of what it did, what happens to software, maintenance, teams and the price of products?

Reviewed · version 3 · 11 min read

Why do tests become the main way to tell an agent what you want, and what makes a test useful to a machine?

Reviewed · version 4 · 11 min read

When agents produce most of the code, what should humans review, and how do teams avoid a review bottleneck?

Reviewed · version 3 · 11 min read

If the context window holds a million tokens, why does the model still miss things?

Reviewed · version 3 · 11 min read

Where does an AI system keep what it knows, and why does it recall the wrong thing?

Reviewed · version 3 · 11 min read

How should knowledge be written when the main reader is a model that keeps improving?

Reviewed · version 3 · 11 min read

Why does one model feel brilliant in one product and ordinary in another?

Reviewed · version 3 · 10 min read

How do you let an agent act without letting anyone who writes text control it?

Reviewed · version 3 · 11 min read

When should steps be fixed in code and when should a model decide what to do next?

Reviewed · version 3 · 10 min read

How do retries, idempotency and uncertain states keep an agent from paying twice or sending the same email three times?

Reviewed · version 3 · 11 min read

How should autonomy change depending on whether an action can be undone, compensated or neither?

Reviewed · version 3 · 11 min read

What problem did the Model Context Protocol solve, why did competitors adopt it, and what does it not solve?

Reviewed · version 3 · 10 min read

What can a ranking actually tell you about which model to use?

Reviewed · version 4 · 11 min read

How do you turn past cases into a test that tells you whether a model, prompt or tool is better for your work?

Reviewed · version 2 · 10 min read

Prices per token keep falling; why do AI bills keep rising, and where will budgets move?

Reviewed · version 3 · 10 min read

How do inference costs change the economics of software companies built on models?

Reviewed · version 3 · 11 min read

Why do incumbents with mediocre AI often beat startups with better AI, and when don't they?

Reviewed · version 3 · 11 min read

How do you find products a new capability makes possible instead of bolting text generation onto every screen?

Reviewed · version 3 · 11 min read

Which products built on someone else's model get stronger as models improve, which get absorbed, and how do you design for the first?

Reviewed · version 3 · 9 min read

Why does a demo that works become a department that doesn't, and what does it take to close the gap?

Reviewed · version 3 · 11 min read

How do you add AI to an organization without a migration nobody wants?

Reviewed · version 3 · 8 min read

Why do permissions, security reviews and data policies decide AI projects more often than models do, and how do you get through them?

Reviewed · version 3 · 11 min read

What job does a process, approval or role do before you automate or remove it?

Reviewed · version 3 · 9 min read

Why can one person get much faster with AI while the team barely changes?

Reviewed · version 3 · 10 min read

How should a company sequence, staff, fund and measure an AI transformation so it produces results in months, not slides?

Reviewed · version 3 · 11 min read

What do forward-deployed engineers actually do, which parts of the work should they own, and how do you avoid making the customer permanently dependent on them?

Reviewed · version 3 · 10 min read

How do roles, review, planning and team size change when each person directs several agents?

Reviewed · version 3 · 11 min read

AI made a step ten times faster. Why didn't the job get faster, and where did the time go?

Reviewed · version 3 · 9 min read

What does a project leave behind that makes the next one cheaper, and what makes AI work repeat itself?

Reviewed · version 2 · 11 min read

Which metrics show real value (lead time, quality, cost per finished task, counterfactuals) and which only look good?

Reviewed · version 3 · 10 min read

What is the difference between producing more, achieving a result, and that result being worth something to someone?

Reviewed · version 3 · 10 min read

How is AI's value split between customers, employees, vendors and owners, and what decides the split?

Reviewed · version 2 · 11 min read

When options change every few months, when should you build, buy, integrate or hire?

Reviewed · version 2 · 11 min read

Which budget lines fund AI adoption, and how does the purchase change when it competes with software, consultants or headcount?

Reviewed · version 3 · 10 min read

Does it matter whether a model's weights are public: control, cost, residency, incentives?

Reviewed · version 3 · 11 min read

Why are model labs moving into agents, apps and enterprise sales, and what does that mean for everyone building on them?

Reviewed · version 3 · 9 min read

What separates a technical edge from a useful product and from an economically sustainable company?

Reviewed · version 3 · 11 min read

If software gets cheap to produce, what stays hard to copy?

Reviewed · version 3 · 10 min read

Why do large companies pay what looks like too much for small AI teams, and when is it a mistake?

Reviewed · version 2 · 11 min read

How do you capture tacit know-how (exceptions, judgment, the why) so people and models can use it?

Reviewed · version 3 · 11 min read

Why was support the first department to change, what the evidence shows, and what other departments can learn from it?

Reviewed · version 3 · 11 min read

What happens to consultancies, agencies and software services firms when AI cuts the hours a job takes?

Reviewed · version 3 · 10 min read

How do accounting and finance use AI when every number must reconcile and be auditable?

Reviewed · version 3 · 11 min read

What happens to publishers, search and online marketing when people get answers instead of clicks?

Reviewed · version 3 · 10 min read

I'm new to this and drowning in terms. Where do I start, and how do I get good?

Reviewed · version 2 · 10 min read

What drives resistance (identity, trust, workload, fear), which parts are signals worth listening to, and how do teams respond?

Reviewed · version 3 · 8 min read

What does research say about deskilling and over-reliance, and how do you keep the judgment you need to check the machine?

Reviewed · version 3 · 9 min read

What does automating part of a job tell you about employment, demand and organization, and what doesn't it tell you?

Reviewed · version 3 · 11 min read

What does the EU AI Act actually require, of whom, and from when, and where do its lines fall?

Reviewed · version 3 · 11 min read

Where should you begin, given what you need?

Reviewed · version 2 · 5 min read

What is actually changing with AI, and where does that leave a person, a team or a company?

Reviewed · version 3 · 14 min read

One verified recent event explained through a mechanism, dated and linked to its durable piece.

Reviewed · version 2 · 7 min read