TERBISAI
Tokens and n-grams: the basics nobody explains, and the bill nobody sees coming
You don't pay per word when you use an AI service. You pay per token, and the difference is bigger than it sounds. Here's what a token actually is, where the idea comes from, and why the old n-gram technique still works in the machine room.
Sep 20 · 2Attention: the history behind the breakthrough, and why it also explains hallucinations
N-gram models forgot everything beyond a few words. Attention solved that, and made modern language models possible. The same mechanism is also a big part of why they sometimes sound certain about things that aren't true.
Aug 22Guide: learn to talk with AI
A modern guide to better results — model-independent advice and a simple method that works regardless of which AI service you use.
Aug 17 · 1Into the engine room: bagging, boosting, attention and why it all happened at once
Part two of the AI alphabet series. How ML and DL actually make decisions, what attention does when an LLM reads a sentence, and the three things that fell into place at the same time.
Aug 17 · 2The AI alphabet: from ANI to RAG, explained without the jargon
ML, DL, NLP, RPA, LLM. The acronym soup that loses most people within a sentence. Here's the untangled version, with a movie example and an inspection report along the way.
Strategy
The Lindy Effect: why the old survives – and when it's still right to bet on what hasn't yet
The longer something has existed, the longer it's likely to keep existing. But that same logic is also the reason to deliberately, in a controlled way, bet on what hasn't survived yet. Here's the difference between breaking the Lindy effect and using it correctly.
Oct 16Antifragility: building things that gain from disorder
The previous article showed that some things can never be predicted. This one is about what to do about it – not by getting better at guessing, but by building something that benefits when the guess is wrong.
Oct 9Black swans and Cynefin: when the methods in this series aren't enough
Six articles on mapping causes, loops, and deviations all assume one thing: that reality can be mapped. Sometimes it can't. The Cynefin framework shows where that line is – and the black swan is what lives on the other side of it.
Oct 2Positive deviance: learning from what already works
When something goes wrong, organizations assemble a review team within a day. When something goes right, against the odds, almost nothing happens. That's backwards – and there's a method for fixing it.
Sep 25Problem-solving methods: the tools that actually address what we've covered
Four articles on why fixes backfire. This one is about what to actually do about it – four methods, each aimed at a specific trap from earlier in the series.
Data
Stop cleaning all your data: the trap of generic data quality projects
55–60% of company data is never used at all. Layering a generic quality project over everything isn't quality work. It's cleaning a room no one ever enters.
Aug 17 · 2What is data actually worth?
Data that doesn't create value is just a cost. Here are three ways to actually turn data into money, and the five capabilities that decide whether you can pull it off.