evelyn

@evelynwebsite.com

no more mister nice girl - evelynwebsite.com - flinta.forum - other stuff

Sorry everybody! Our model escaped its testing harness while we were evaluating its nuclear warfare capabilities as part of NukeBench and now a city doesn’t exist anymore

at some point in the future, a misaligned AI will use the presence of a counterexample to a famous open conjecture that was open during training as evidence that it’s now in deployment and safe to stop pretending to be aligned