#ai reels — Knowledge Reels

Swipe up

The test for machine thinking is 75 years old

In 1950 Alan Turing proposed the imitation game: if a person cannot tell a machine from a human in conversation, asking whether it thinks may be the wrong question.

#ai #history #turing

The first neural network ran in 1958

Frank Rosenblatt's Perceptron was a physical machine with motor-driven weights that learned to classify simple shapes — decades before the deep learning boom.

#ai #history #neuralnetworks

AI learns by running its mistakes backwards

Backpropagation sends the error at the output back through the network, nudging millions of weights slightly in the direction that reduces it. Repeat billions of times.

#ai #learning #math

One 2017 paper reshaped modern AI

"Attention Is All You Need" introduced the Transformer, letting models weigh every word against every other word at once. Almost every large model today descends from it.

#ai #transformers #history

Move 37 stunned the world of Go

In 2016 AlphaGo played a move human experts called a mistake. It won the game — and changed how professionals understood a 2,500-year-old game.

#ai #alphago #games

Modern AI is trained on graphics chips

GPUs were built to render game pixels in parallel — the same maths as multiplying huge matrices, which is exactly what training a neural network needs.

#ai #hardware #gpu

Language models do not read words

Text is split into tokens — chunks of characters. A rare word may cost several tokens, which is why models sometimes miscount the letters inside one.

#ai #language #models

AI hallucinations are confident guesses

A language model predicts likely next tokens, not verified truth. When knowledge runs out, fluent invention looks exactly like fluent fact.

#ai #safety #models

A photo dataset kicked off the deep learning era

ImageNet gathered millions of labelled images, and in 2012 a deep network called AlexNet cut the error rate so sharply that the field changed direction overnight.

#ai #history #vision

Human feedback is what makes chatbots helpful

RLHF trains a reward model on human preferences between answers, then tunes the AI to score higher — turning raw prediction into something worth talking to.

#ai #rlhf #training

Reels tagged #ai