I am a PhD researcher working on AI in education. This is where I write down what I read, while it is still untidy.
A model can be 95% right on day one, but if nobody opens it, that figure produces no value at all.
This is a reminder I hold on to: right today does not mean right forever.
This is the gap where I see a great many AI projects stop.
This is a lesson in cost discipline that I see many teams learn a little late.
This is a principle I believe holds for AI and for operations alike: what you cannot see, you cannot manage.
This is the step that turns AI from something that only talks into something that actually does the work.
This is a habit I learned: to find out how solid something is, try breaking it yourself before anyone outside gets the chance.
This is what I believe separates an AI system worth using from one that will cause trouble.
This is the aspect I think will matter more and more as AI moves into tightly regulated industries.
This is what I say whenever someone tells me proudly that their company collects a great deal of data.
This is the question I ask before trusting any AI tool: on what basis do we say it is good?
This is the boundary I believe matters most for using AI that is both powerful and easy to live with.
This is the direction that made me think hardest about how a lean team ought to be organised.
This is the concept that kept me level-headed while the noise about AI was at its loudest.
The deeper I research, the more I believe using AI well and using AI safely have to travel together, especially as the rules tighten.
This is the risk I raise first whenever someone is about to hand important work to AI without checking it.
This is why AI assistants today understand what you mean far better than a few years ago, and not simply because they are bigger.
One figure about cost made me reconsider the line leave AI until next year.
This is the concept that convinced me small enough to do the job is often wiser than large and cumbersome.
This is a fine example of how being cleverer sometimes means doing less, in the right place.
This is the most fundamental cure for the chatbot that keeps making things up, which many people hit when bringing AI into their work.
One research result has stayed with me, because it proves that how we ask matters no less than the model itself.
This is the question I am asked most when a business starts using AI: do we need to retrain a model of our own? In most cases, the answer is not yet.
Every time an AI image turns out unexpectedly beautiful, I remember it was not painted the way we imagine. It was lifted out of a mess.
You hand a very capable person your company's entire file and ask them to find one detail in the middle. They read all of it, and remember the beginning and the end best.
Something small throws most people the first time they use AI: a model that can draft a whole contract miscounts the letters in a simple word.
Have you ever typed a vague search into Google, the keywords all wrong, and it still understood and gave you exactly what you needed?
Nearly every AI tool we use daily stands on one idea, published in 2017 by a group at Google under a very modest title.
In the early days of AI, every task needed its own model built almost from zero: one to translate, one to recognise images, another to filter spam.
The field has a stock phrase so blunt it sounds like a joke: garbage in, garbage out.
Something troubles the AI field: a model can predict very accurately and still not say why it concluded what it did.
You have probably felt this: a few minutes into TikTok or Netflix and it has already read you, serving exactly what you wanted, though you never said so.
One of the great turns in technology was the day computers began to see. Before that, teaching a machine to recognise a cat meant describing every feature by hand.
In teaching machines there is a small distinction I find illuminating about management: between what the machine learns itself, and the few things a person must set by hand beforehand.
Many people assume the machine reads through the data once and is done. The reality is the opposite: it studies the same material over and over.
Remember the game of telephone? A sentence whispered through seven people, and by the last one it has gone crooked. Deep networks hit exactly that problem.
Why do some models get past a hard patch easily while others get stuck halfway? Often the difference is something very familiar: momentum.
What decides whether a cell in a neural network should speak up or stay quiet? The answer says a good deal about how we decide things.
There is a trick that makes a model stronger and sounds absurd at first: they deliberately switch part of it off.
We have all met the student who knows every exercise in the book by heart and freezes at a question phrased slightly differently.
When teaching a model, one tiny dial decides almost the whole result: after each lesson, how much should it change itself?
It sounds backwards, and it is the heart of how AI learns: what makes it better is not the times it guessed right, but the times it guessed wrong.
Think back to how a child learns to recognise digits. Nobody reads them a definition. They look, guess, get it wrong, get corrected, and gradually know.
Mỗi sáng tôi gửi một việc nhỏ bạn làm được ngay hôm đó với AI. Đọc hết trong hai phút.