GRAM could let one model be trained once and deployed with or without dangerous dual-use knowledge, rather than training separate filtered models. It is preliminary, so treat it as a research direction.
Key takeaways · AI-distilled
GRAM (Gradient-Routed Auxiliary Modules) adds extra neurons to every transformerThe neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.Full definition → layer, grouped into one module per dual-use category. On dual-use text only that category's module learns while the general weights are temporarily frozen.
After training, a module can be deleted to remove its capability or kept for trusted deployments. With four categories, one training run yields a model configurable 16 ways, where data filtering would need separate models.
In a larger test with virology, cybersecurity, nuclear physics and a niche programming language as dual-use domains, deleting a module removed that capability about as effectively as never training on the data.
Anthropic and AE Studio call the results preliminary: GRAM has not been applied to any production Anthropic model, and the post says the team is not sure it ever will be.
Terms in this piece · Glossary
transformer — The neural network architecture behind modern AI models, built on attention — letting every word directly consider every other word in parallel.
