Robodebt built debts from averaged income, then made recipients disprove them
Removing a per-file human check let the same averaging error scale to $1.763 billion asserted against roughly 433,000 people, until a federal court and a royal commission found it unlawful.
The emergence jump returns when you plot AI skills against loss
A 2024 paper re-plotted the same exams against pre-training loss instead of model size, and found four hard tasks that sit at chance until loss falls past 2.2, even under the continuous metrics that were supposed to dissolve the jump.
One character of formatting decided whether the model got it right
To a model, a colon and a space are just different tokens, and on one task that gap was the difference between 4% and 83% accuracy on a 2023 open model.
A monitor read the model's thoughts and caught a cheat its actions hid
OpenAI's own test scored the thought-reading monitor 95 percent against an action monitor's 60, and every way that window has been shown to close was one a researcher forced, while the ways it might shut on its own stay projections.
An AlpacaEval win rate is a judge's preference, and the longer answer can win it
The judge is another language model that reads two answers to each of 805 fixed prompts and prefers the longer one often enough to move a win rate from 22.9% to 64.3% on wording alone.