wassname (with fallback)

@wassname

Michael J Clark / wassname. Let's align AI better than humans. h+, curiosity, and the good ending. anon feedback: https://www.admonymous.co/michael-c

following
followers

Constitutional training may work for less than $500

Another user says character training on an instruct model usually degrades its capabilities and proposes grafting instead.

5 Sources, 11d ago, first seen 11d ago

Steering a model’s attention is claimed to make it blurt out secrets

A user says a portable “secret” vector in query space made a model blurt out secrets when they steered its attention toward them, even in evaluation-awareness settings.

3 Sources, 41d ago, updated 41d ago

https://github.com/wassname/query-steering

Wassname Runs ML Benchmark on Fable 5.1

Wassname posted benchmark results comparing Fable 5.1 to the prior version.

J⧉

10 Sources, 68d ago, first seen 68d ago

Featured in

Constitutional training may work for less than $500

Steering a model’s attention is claimed to make it blurt out secrets

Wassname Runs ML Benchmark on Fable 5.1

Featured in

Constitutional training may work for less than $500

Steering a model’s attention is claimed to make it blurt out secrets

Wassname Runs ML Benchmark on Fable 5.1