The AI Reader.
Saturday, August 15, 202647 min read
Working Knowledge · What Could Go Wrong

Bostrom says only luck has kept us from a world-ending technology

The cure he proposes is a tag around every neck that uploads your whole day for a machine to watch, and even readers who buy the argument balk at it.

10 min read8 sources
Working Knowledge · When AI Breaks

To stop labeling Black men as gorillas, Google stopped labeling gorillas

In 2015, Google Photos filed pictures of Jacky Alciné and a friend, both Black, into an album it labeled Gorillas, and eight years later that same block was still the best answer four companies had.

10 min read8 sources
Working Knowledge · The Evidence

The knowledge distillation paper trained a network to recognize a digit it never saw

Hinton, Vinyals, and Dean's 2015 paper trained the small network on how the big one ranked every wrong answer, and that soft-target method is the part much of today's 'distillation' leaves out.

9 min read7 sources
Working Knowledge · The Instruments

HellaSwag's right answer is just the sentence that came next

The 2019 test built to stump machines is now solved, and much of its answer key turns out to be wrong.

9 min read9 sources
Working Knowledge · The Mechanics

A repeated phrase makes a model likelier to repeat it again

The loop shows up under greedy and beam search but not under sampling, and why a trained model favors its own recent words in the first place is still unsettled.

9 min read8 sources