The AI Reader.
Friday, August 7, 202646 min read
Working Knowledge · What Could Go Wrong

Soares warns that AI safety could break the moment capability jumps

The measured gap behind the fear is an arcade agent that quits chasing a coin, and the sudden system-wide turn it warns of is still an analogy to how evolution built humans.

10 min read9 sources
Working Knowledge · The Evidence

The deeper network fit its training data worse

Microsoft's ResNet swept the 2015 vision contests by adding each layer's input back to its output, the skip connection now wired into every transformer.

9 min read6 sources
Working Knowledge · The Instruments

A faithful summary of a false passage passes the hallucination test

Vectara's leaderboard, the number behind every 'which AI hallucinates least' ranking, scores each model on how faithfully it summarizes a passage it was handed.

9 min read9 sources
Working Knowledge · The Mechanics

A transformer can only compute more by writing more

A fixed network does a bounded amount of computation per token, so the only way to spend more on a hard problem is to write the intermediate steps and read them back.

9 min read12 sources
Working Knowledge · When AI Breaks

New York cleared the Apple Card of bias without proving it fair

In 2019 David Heinemeier Hansson shared every bank account with his wife and was offered twenty times her Apple Card limit, and Goldman Sachs insisted its model never knew which of them was a woman.

9 min read8 sources