Blog

What the “don’t train on my data” toggle does not do

Your “Don’t Train on My Data” Toggle Does Not Do What You Think

The toggle answers exactly one question. It says nothing about where your data is processed, how long it is kept,

October 20, 2026

Self-hosting is a utilisation question wearing a privacy costume

Self-Hosting Is a Utilisation Question Wearing a Privacy Costume

The business case for self-hosting an LLM is usually argued on privacy and justified on cost per token. Both are

October 5, 2026

Nobody else’s similarity threshold is right for your data

Nobody Else’s Similarity Threshold Is Right for Your Data

A cosine score of 0.75 has no universal meaning. If your relevance threshold was copied from documentation rather than derived

September 21, 2026

Why the cheapest capable AI models are open weight

Nobody Is Being Generous: Why the Cheapest Capable AI Models Are Open Weight

Several Chinese laboratories now publish models close enough to the frontier to make a genuine choice possible. The reasons are

September 10, 2026

Reasoning mode is making your AI worse at grounded tasks

Reasoning Mode Is Making Your AI Worse at the Thing You Are Using It For

On grounded tasks, several frontier reasoning variants hallucinate more than their non-reasoning counterparts. Thinking harder and staying faithful to a

August 21, 2026

Most People Stop at Mount Stupid With AI. The Risk Is Never Going Past It.

Most People Stop at Mount Stupid With AI. The Risk Is Never Going Past It.

When you first use AI you feel like an expert at everything, that's Mount Stupid. The more you use it,

July 13, 2026

We Were All Farmers Once. Nobody Calls That a Tragedy.

We Were All Farmers Once. Nobody Calls That a Tragedy.

Junior skills will change; spreadsheets and contracts will increasingly write themselves. But entry-level work evolving is not the same as

July 10, 2026

The Discount Reflex Isn't a Pricing Problem. It's a Knowledge Problem.

The Discount Reflex Isn’t a Pricing Problem. It’s a Knowledge Problem.

I signed 7-figure deals with heavy discounts; several customers had budget for 10-15% more. When your best salesperson leaves, the

July 8, 2026

Global Threats Are Accelerating. Defence Engineering Capacity Is Not.

Global Threats Are Accelerating. Defence Engineering Capacity Is Not.

Europe's defence industrial base is short 150,000-200,000 skilled workers; Barrow's submarine workforce fell from 14,000 to under 3,000. You cannot

July 5, 2026

AI Doesn't Have to Burn a Coal Plant to Answer a Support Email

AI Doesn’t Have to Burn a Coal Plant to Answer a Support Email

A 175B cloud model to triage a ticket a 9B model could handle on a single office GPU. We mapped

July 2, 2026