Alignment Forum
A specialized forum for AI safety and alignment research, incubated by the LessWrong community. It focuses on academic discussion and proposal publication in the areas of machine learning alignment and AI safety research.
Timeline (last 90 days)
Fixed-weight models (at least as we understand them today) will always have adversarial fragility in their concept space.
WorkspaceBench is a benchmark specifically designed to evaluate the reading capability of the "global workspace," used to judge whether a tool can restore intermediate variables in the forward pass into readable text.
A proposal on the Alignment Forum argues that a lab should not train the next, more capable generation of models before adequately resolving all major alignment issues in the current generation.
All Facts (3)
WorkspaceBench is a benchmark specifically designed to evaluate the reading capability of the "global workspace," used to judge whether a tool can restore intermediate variables in the forward pass into readable text.
50%UnverifiedA proposal on the Alignment Forum argues that a lab should not train the next, more capable generation of models before adequately resolving all major alignment issues in the current generation.
50%UnverifiedFixed-weight models (at least as we understand them today) will always have adversarial fragility in their concept space.
50%