
3
PUBLIC
How they work
Large language models
• It takes some words as input. It generates the next word. It repeats this to make sentences and paragraphs.
• It remembers nothing. It learns nothing. So we need to give it everything it needs to know as input.
• It tries to guess what "sounds" best - not what is true.
• it is able to guess based not just on stats, but by generalising, "reasoning", and other things
function guessNextWord( input: words Array<String> ) output: String {
var numbers = convertWordsToNumbers(words);
var nextNumber = guessNextNumber(numbers);
return convertNumberToWord(nextNumber);
}
function guessNextNumber( input: numbers Array<Number> ) output: Number {
if (numbers.size > 8192) throw "Sorry! Too many inputs";
var guess = numbers[0] * 0.123124 + numbers[1] * 0.584935 + ...
// Huge amount of adding and multiplying numbers, but no state nor network calls etc
// Training the model takes a huge amount of time and money
return guess;
}
文档被以下合辑收录
相关文档
评论