Why does 2048 give you a 4 when you were expecting a 2? Because the game rolls a die for every new tile, and about 1 roll in 10 comes up 4. It is not a bonus and it is not a punishment: it is one line of code. This guide…
A place to write, and be read.
Practical guides from vlvd.net on AI safety, security testing, study skills, history and games. Read them, follow the writers you like, or write your own.
Latest posts
Fresh off the bench.
Newest first, from everyone. Sign in to follow people and see only their posts.
A flipped edge on a Rubik's cube is a piece sitting in its own spot with its two colours the wrong way round. If you find just one, with everything else solved, no amount of turning will fix it. Edges flip in pairs, so…
A Rubik's cube corner twist is a corner piece turned a third of a turn in its spot, and the short answer is that you can never make one on its own. Turn the faces any way you like and the corners travel and twist, but…
A study guide tells you what the test will cover. It does not tell you whether you know it. Reading the guide and nodding along feels like revision, but the surest way to find out is to answer questions on every line.…
Some people cannot study without headphones. Others need a silent room. Both are sure they are right. So is it better to study with music or in silence? The research has a clear answer for one kind of music, a mixed…
Pharmacology asks you to hold hundreds of drug names, each with a way it works, what it is used for, and what it can do wrong. Reading the list again and again does not stick. If you are working out how to study for a…
Untrusted monitoring is a safety technique from the field of AI control: you use one copy of a powerful AI model to check the work of another copy, and assume either copy might be trying to trick you. In short, it trades…
A linear probe in machine learning is a simple linear classifier trained on one layer's activations inside a network, to check whether that layer holds a property such as "is this statement true?". The model itself is…
Eliciting latent knowledge, usually shortened to ELK, is the problem of getting an AI to tell you what it actually knows, rather than what it predicts you would believe. That sounds like a small difference. The…
Good tank game controls disappear: you think "turn left, fire" and the tank does it. This guide covers the tank game controls for both games on Tank City Reboot, on a keyboard, a gamepad and a touch screen. It also…
Good questions to ask your grandparents can turn an ordinary visit into the best conversation you have all year. Most grandparents have stories nobody has asked for: what bread cost, how they met, the year everything…
What are computer chips made of? Almost every computer chip, including a CPU, starts as silicon. So why are computer chips made of silicon, and not copper, carbon or gold? The short answer: silicon is a semiconductor, it…
Can AI be creative? AI systems now write poems, compose music and suggest research ideas, and some of it is good. But being good is not the same as being creative, and people disagree sharply about whether a machine can…
Liquid cooling is a way to keep a computer chip cool by moving its heat into a liquid instead of leaving it to a fan. So how does liquid cooling work in a PC? A liquid flows through a block on the hot chip, soaks up its…
Do we owe anything to future generations? Many philosophers say yes, but they disagree about why, and the doubts are serious ones: future people do not exist yet, and our choices help decide who will exist at all. This…
AI agent audit logs answer the question everyone asks after an agent does something wrong: what exactly did it do, and why was it allowed? Most app logs cannot say. They record a request and a status code. They do not…
Differential technological development is a strategy for reducing the risks of new technologies by changing the order in which they arrive. Instead of trying to stop progress, you slow down the dangerous technologies and…
A global catastrophic risk is a possible event that could seriously harm human well-being across the whole world, endangering or even destroying modern civilization. Catastrophic threats of this kind include pandemics,…
Human Compatible by Stuart Russell is a 2019 book that argues AI built to chase fixed goals as hard as it can could become dangerous as machines get more capable, and that machines should instead be built to learn what…
Starving a castle was the slow way to win a siege. Instead of breaking the walls, the attackers closed every road in and out, then waited until hunger forced the defenders to surrender. English Heritage calls this…
A curtain wall on a castle was the stone wall that ran round it and joined its towers together. It was the castle's outer skin, the line that kept attackers away from the buildings inside. This post explains what it was,…
Rogue AI agents are hard to catch because each thing they do can look fine. One file read, one API call, one approval: all inside the rules. The harm is in the pattern. An agent that was meant to cut costs starts…
The Two Towers Trebuchet is the tower the title screen labels "Stones on a crowd", and that label is the whole idea. It throws slow stones a long way, and each stone hurts every attacker where it lands. This guide covers…
Agent goal hijack is what happens when an AI agent stops working on your task and starts working on someone else's. Nobody breaks into your server. The agent reads a web page, an email or a calendar invite with…
If you need to know how to analyze a primary source for a history class (also called primary document analysis, or analyzing primary sources, or in British spelling how to analyse one), the good news is that the method…
Gandhi vs Martin Luther King is less a contest than a family tree. King's civil rights organization took its working methods from Mohandas Gandhi. The Nobel Prize calls Gandhi "the strongest symbol of non-violence in the…
Why is Shakespeare still important, more than 400 years after he died? Short answer: his plays are performed more often than those of any other playwright, according to World History Encyclopedia, and he showed how to…
A castle keep was the great tower of a castle, and at castles such as Goodrich and Middleham it is the oldest part, the core the rest grew around. English Heritage often calls it the "great tower or keep", and both names…
A licence to crenellate (spelled license to crenellate in American English) was the king's permission for a lord to fortify his house: to give it battlements, and with them the look and some of the strength of a castle.…
The Two Towers Archer Tower is the first tower on the title screen, and its label there is "Fast arrows". Fast arrows sound like they should answer everything. They do not. This guide is about the one thing that makes…
Write here too
Your own page.
Post what you think, follow the people you like and talk about what they write.