The Reward Model Was Hiding in the Policy
The reward model RLHF pays a full training run to fit is already implied by the policy; Direct Preference Optimization turns that identity into a single classification loss.
The Transform That Was Too Simple to Fund
A grant reviewer dismissed the discrete cosine transform as too simple in 1972; it now rebuilds nearly every image and video frame on Earth, and it works because it cheaply imitates a transform we could not afford to compute.
Was the United States Right to Decline a 16-Year USMCA Renewal?
By refusing to extend USMCA to 2036 in favor of yearly reviews, the Trump administration wagered leverage against certainty, and serious people disagree about which matters more.
A Flat Mile Around Swan Lake
At Rockefeller Preserve, the Rockefellers' crushed-stone carriage roads circle a 24-acre lake on a grade gentle enough to push a newborn around before the July heat lands.
The Frontier Gets Cheaper Faster Than We Can Grade It
OpenAI, xAI and Meta all cut the price of intelligence this week, while two fresh audits show the cheap models we use to grade each other are shakier than the invoice suggests.
A caught rocket, a record-hot June, and Russia's fuel squeeze
China nets an orbital booster and Europe logs its hottest June while Ukrainian strikes ration Russia's petrol and Ebola spreads unseen in Congo.