2017–2026
40 entries
A living archive of thoughts.
Notes, essays, research seeds and reading notes, numbered in the order I wrote them.
Imperfect and evolving.
- N° 40
Garbage In: The Unglamorous Half of Machine Learning
Fifty rows of random numbers, labels with nothing to do with them, and a model that scores 95 per cent. A tour of the ways data misleads a model before it is trained, and the one mistake behind that number.
notemachine learning, data, statistics
- N° 39
Writing Knowledge Down
Tweety is a bird, so Tweety flies. Opus is a bird too. What happens when you try to give a machine a rule and an exception, why a graph of facts can resolve which Apple you meant, and the price of a language that can say anything.
noteai, knowledge representation, history of ai
- N° 38
Finding the Needle: How Retrieval Ranks Documents
A search engine does not read your documents when you ask. It read them once, in advance, and filed them under every word. Then comes the harder part: ranking what it finds, and checking that the score you trust is really a score.
notenlp, search, python
- N° 37
How a Convolutional Network Sees
A plain network and a convolutional one read the same handwritten digits, and one of them needs 127 times fewer adjustable numbers. Where the saving comes from, checked by hand, and what happens when you borrow, stretch or forge pictures to train with.
noteai, computer vision, neural networks
- N° 36
The Naive Assumption That Works Anyway
Naive Bayes multiplies probabilities as if the words in a review had nothing to do with each other. They plainly do. Why that false assumption still picks the right answer so often, what one zero count can do to it, and the small experiment where it finally breaks.
noteai, probability, machine learning
- N° 35
How Machines Learn to Read
A person reads a sentence in a second. A machine has to cut it, count it, place it in space and weigh it, and every step either fixes a problem left by the last one or deletes something you needed. A walk through the stages, with code you can run.
noteai, nlp, machine learning
- N° 34
Memorising Is Not Learning
A curve that passes through all twelve points exactly, with a training error of 0.0000, and an error of 5.89 on new ones. Overfitting, the two ways a model can be wrong, why the test set is honest only once, and a perfect score earned on pure noise.
notemachine learning, statistics, data
- N° 33
Backpropagation: Learning by Assigning Blame
The sentence I first met was that the error is fed back and the weights are adjusted. Here is what that actually means: one tiny network, every number worked out and checked, and why the blame has to be shared out layer by layer.
noteai, neural networks, optimisation
- N° 32
One Neuron, One Line, and the Problem That Froze a Field
Four points on a square and a rule a child can state, which a single artificial neuron can never learn. Why it cannot, what one extra layer changes, and what the popular story about a frozen field gets right and wrong.
noteai, neural networks, history of ai
- N° 31
Evolution as an Algorithm
A genetic algorithm found the best route through six cities in a blink, and it was the right answer to the wrong question. How a search that copies evolution works, what its three dials really do, and why it can throw away its best answer and still win.
noteai, optimisation, search
- N° 30
What a p-value Is Not
A machine that never finds a real effect still announces a discovery about once in twenty runs. Running it 10,000 times is the quickest way I know to see what p < 0.05 does and does not say.
notestatistics, probability, python
- N° 29
The Shape of Data
Four datasets share a mean, a spread, a correlation and a fitted line, and look nothing alike. A walk through the numbers we use to summarise data, what each one hides, and the question I should have asked before any of them.
notestatistics, data, probability
- N° 28
Why Data Is a Matrix
A photograph, a customer list and three film reviews look nothing alike, yet inside a model they are the same object. How a grid of numbers turns similarity into angles, a layer into one product, and a table into something that can act.
notemathematics, machine learning, data
- N° 27
Drawing Uncertainty: Bayesian Networks
Five yes-or-no facts about a house need 31 numbers to describe completely. Draw the arrows between them honestly and ten will do. A runnable burglar alarm, and what a neighbour's phone call is really worth.
noteai, probability, python
- N° 26
Bayes' Rule, or How to Change Your Mind
A test that catches 80 per cent of cases comes back positive, and your chance of being ill is about 7.5 per cent. The gap is the whole of Bayes' rule: how to let evidence move a belief without letting it replace the belief.
noteai, probability, mathematics
- N° 25
A*, and the Art of the Honest Guess
Four tiles are out of place, so the puzzle takes four moves. It does not, and the gap between that quick count and the truth is the whole art of A*. One honest guess saves the search from wandering, and a boastful one quietly breaks it.
noteai, search, python
- N° 24
How a Machine Searches
Take the oldest item off the waiting list and you get ripples. Take the newest and you get a hiker who never turns back. One small swap, opposite behaviour, and why the route with fewer turns is not always the cheaper one.
noteai, search, problem framing
- N° 23
Why Python Won AI
The language the whole field runs on is one of the slow ones. I timed it, and then followed the puzzle down to what Python is really doing: calling other people's fast code, behind an interface that makes trying the next idea cheap.
notepython, ai, machine learning
- N° 22
An Agent and Its World
A wasp with a flawless routine and no way to notice that the world has moved. What it takes for a machine to do better: what rational really means, how to describe the world it lives in, and a ladder of designs where each rung fixes the one below.
noteai, problem framing, rational agents
- N° 21
The Off-Switch and the Objective: What the Fable 5 Shutdown Reveals About Capability Control
On 12 June 2026, the US government ordered Anthropic to disable its most capable public model. The action was not a response to a misaligned system. It was the off-switch being pulled, and it exposes the precise limit that Bostrom and Russell warned about: capability control is a supplement to alignment, never a substitute for it.
essayai, ai safety, value alignment
- N° 20
What Are We Actually Trying to Build?
Turing asked whether machines can think, then refused to answer his own question. The field that followed has four different ideas of what it is building, and the one it chose has a crack running right through it.
noteai, history of ai, problem framing
- N° 19
He Asked for Water. His Son Threw Him in the Sea. This is an AI Problem.
What a Arabic proverb and a drowning man can teach us about how we state problems in AI. The right question is half the answer. The wrong goal is all the danger.
essayai, problem framing, requirements
- N° 18
husainalghasra.com
My personal website and public notebook. Built with Gatsby, React, and Markdown. A living knowledge system.
buildgatsby, react, markdown
- N° 17
AI Traceability: The Gap Between Answer and Trust
AI systems can produce confident answers with no trail of reasoning behind them. This is not just a UX problem. It is a fundamental problem in epistemics and system design.
research seedai, machine learning, traceability
- N° 15
Data Is Not Information
Two bytes can be a word, two different numbers, a pair of grey pixels or a Chinese character. None of those meanings is in the bytes. Where information actually lives, and why Shannon threw meaning out on purpose.
notedata, information theory, encoding
- N° 14
The Twitter Data Breach: What You Need to Know and How to Protect Yourself
200 million email addresses exposed. What the Twitter breach tells us about API security, data minimisation, and why email addresses are more dangerous than people think.
articlecybersecurity, data breach, twitter
- N° 13
Frames of Mind: The Theory of Multiple Intelligences by Howard Gardner (Part 2)
Continuing through Gardner's intelligences: bodily-kinaesthetic, musical, spatial, intrapersonal, interpersonal. What the framework gets right and where it gets contested.
articleframes of mind, multiple intelligences, howard gardner
- N° 12
Frames of Mind: The Theory of Multiple Intelligences by Howard Gardner (Part 1)
Gardner's challenge to the IQ monoculture. A still-contested framework that asks whether linguistic-logical intelligence is the only kind worth measuring.
articlebrain, education, frames of mind
- N° 11
CoinMarketCap Data Breach
3.1 million user emails compromised. A breakdown of what happened, what it means, and what you should do if your email was in the breach.
articleblockchain, coinmarketcap, cyber security
- N° 10
The New One Minute Manager
Simple management principles wrapped in a fable. The ideas are obvious once stated, but that is the point.
noteken blanchard, leadership, management
- N° 9
What Start With Why Actually Changed in How I Think
Sinek's book is easy to dismiss as business self-help. But asking why before what changed how I approach problems in ways I did not expect.
essaybusiness, leadership, start with why
- N° 8
Start With Why: Simon Sinek
The Golden Circle is a genuinely useful mental model even if the book overstates its case. Worth reading for the core idea.
notebusiness, book reviews, simon sinek
- N° 7
Exploring the Strange Beauty of The Appointment by Herta Müller
Müller writes under surveillance, both literal and psychological. Her prose is fractured and precise in a way that matches the subject perfectly.
notebook reviews, novels, herta muller
- N° 6
Diving into Philosophy: My Experience with Josiah Royce's The World and the Individual
Royce's idealist philosophy is unfashionable and demanding. But thinking through why he was wrong turns out to be useful.
notethe world and the individual, existence, josiah royce
- N° 5
What is Relativity? by François Vannucci: A Disappointing Exploration
A book that promises an accessible introduction to relativity and delivers a poor translation. A useful lesson in what happens when precision in language is missing.
noteeinstein, relativity theory, science
- N° 4
A Review of Listen, Little Man! by Wilhelm Reich
Reich's passionate critique of mass psychology and self-inflicted oppression. A short, uncomfortable, and necessary book.
notephilosophy, book reviews, wilhelm reich
- N° 3
The Thirst for Freedom: A Review of Emmi Itäranta’s Memory of Water
A quiet, haunting novel set in a world where water is power. Itäranta builds a future that feels uncomfortably close.
notebook reviews, novels, emmi itaranta
- N° 2
Understanding Modern Politics through Giorgio Agamben's State of Exception
Agamben argues that emergency powers are not a temporary suspension of the law but a permanent feature of modern governance. Dense and important.
notestate of exception, giorgio agamben, politics
- N° 1
A Thousand Splendid Suns: A Story of Love, Loss, and Survival in War-Torn Afghanistan
Hosseini's second novel is relentless and moving. A story of two women, decades of war, and the weight of staying alive.
notebook reviews, novels, khaled hosseini