A physicist proposes a test: train an AI on physics only up to 1925 and see whether it invents quantum mechanics. If it can, it might find theories we can't guess.
Read more →
AI may use up all public human text by 2032. Sergey Markov offers a half-joke: maybe our world is a generator of synthetic training data for some model.
Read more →
A smartphone can hold every grandmaster chess game ever played. A psychologist argues that human limits, not limitless knowledge, are where creativity begins.
Read more →
The original definition of AGI can't be tested, so companies wrote their own. In one version, AGI is whatever earns a billion dollars. Here is why that matters.
Read more →
A model that can take a million tokens may use only 5–10% of them well. Mikhail Burtsev on why attention blurs, and why a fresh chat often beats a long one.
Read more →
Credit and job decisions now come from algorithms with the weight of a court ruling. Sergey Markov on why you can't see the evidence or appeal.
Read more →
A human genome cut into ten-letter words runs to 300 million tokens. How a language-model trick lets AI read DNA, and why drug doses and gene therapy are the targets.
Read more →
A language model sees only words, yet to guess the next one it may have to rebuild the world behind them. An AI researcher on why that dumb task asks so much.
Read more →
Chatbots get tangled in long chats because attention stretches thin. A working-memory token lets a transformer read in small chunks and carry the thread to 50 million tokens.
Read more →