Telangana at a glance

The smallest unit of information — a single value that is either 0 or 1. Everything a computer stores is built from them.

Eight.

A finite, unambiguous sequence of steps that turns an input into a result. It has nothing to do with computers in principle — long division is an algorithm.

They nest. AI is the broad goal of getting machines to do things that seem to need intelligence. Machine learning is one way of doing it — learning patterns from data instead of being told rules. Deep learning is machine learning using neural networks with many layers.

Adjusting the numbers inside it, over and over, so that its output moves closer to the examples it is being shown. Nothing is being taught in the human sense; numbers are being fitted.

One of the numbers inside a model that training adjusts. A model with more parameters can represent more complicated patterns — it does not follow that it is more truthful.

Supervised learning uses examples that already carry the right answer (a labelled photo, a graded essay). Unsupervised learning is given data with no answers and looks for structure in it.

Layers of very simple units. Each unit adds up its inputs with different weights, applies a simple function, and passes the result on. The behaviour comes from the arrangement and the weights, not from any single unit.

The chunk of text the model actually reads and writes. A token can be a whole word, part of a word, a single character, or a space — it depends on the tokenizer, not on grammar.

By predicting the next token, adding it, and repeating with the longer text as input. A sentence is the result of doing that many times, not of planning a sentence.

Because the mechanism optimises for a plausible continuation, not a true one. Nothing in the process checks a claim against the world, so a fluent wrong answer and a fluent right answer are produced the same way. Confidence in the wording is not evidence.

Not by itself. It has no separate store of verified facts to check against, so it cannot reliably mark its own output as right or wrong. Verification has to come from outside it.

Because a reference that looks correct — plausible author, plausible year, plausible title — can be generated without existing. The only test is opening the source. This site has published a correction for exactly that kind of mistake.

When a model fits its training data so closely that it has partly memorised it, and does worse on data it has not seen. Doing well on material you studied is not the same as understanding it.

Because training is overwhelmingly matrix multiplication — huge numbers of identical simple operations. A GPU runs many of them at once, where a CPU is built to run fewer, more varied ones quickly.

A systematic skew in its output — one that leans consistently in a direction — usually inherited from the data it was trained on or the choices made in building it. It is a property of the system, not an opinion the system holds.

That no amount of processing rescues bad input. A model trained on unreliable data produces unreliable output, however sophisticated the method.

The collection of examples a model is trained or tested on. What is in it — and what is missing from it — shapes the result more than most other choices.

An agreed way for one program to ask another for something: what you may ask, how to ask it, and what comes back. It lets programs work together without either knowing how the other is built.

Software whose source code is published under a licence that lets others read, change and share it. Open source is about the licence and the code, not about price.

Lossless compression can rebuild the original exactly. Lossy compression throws away detail judged less noticeable, so the original cannot be recovered — which is why repeatedly re-saving a photo degrades it.

A test proposed by Alan Turing in 1950, in his paper Computing Machinery and Intelligence. He suggested replacing “can machines think?” with a practical question: whether a person questioning a machine and a person could tell which was which.

Gordon Moore's 1965 observation that the number of transistors on a chip was doubling at a regular interval — originally about every year, later revised to roughly every two. It is a trend in an industry, not a law of physics.

Because most tokenizers were built from text that is overwhelmingly English, so English words often survive as one token while Telugu gets split into many smaller pieces. Since cost and context limits are counted in tokens, the same meaning is more expensive to process in Telugu. Published measurements of the size of that gap vary, so we do not print a single figure.