This is a targeted grant for a specific set of data to generate for Decode Research, on Neuronpedia. These are "feature dashboards" and autointerp for a new set of SAEs that we wish to generate. The amount has been pre-agreed upon.
This is for a specific compute-heavy task called "feature dashboards" for Decode Research.
We have already generated about 20% of these dashboards, an example of a dashboard is: http://neuronpedia.org/gemma-2-2b/0-gemmascope-res-16k/34
Here's a listing of each SAE that will have dashboards (some are still hidden for now):
https://www.neuronpedia.org/gemma-scope#browse
The overall Gemma Scope project on Neuronpedia is here: https://www.neuronpedia.org/gemma-scope#main
We plan to, in all, generate 40,000,000 dashboards and do auto-interp for them too. Auto-interp is basically asking an AI like GPT4 to interpret what a feature is about.
We will spend all of it on either feature dashboards or auto-interp explanations. We expect to be cost-neutral on this - eg it goes entirely to compute.
Joseph Bloom, Curt Tigges, Johnny Lin of Decode Research. We have generated ~20% of the dashboards already.
This will likely not fail unless something really weird happens, like all AI compute resources are suddenly 10x the expected price.
This is a targeted grant for a specific task. Our project has gotten other funds from various sources, but none of it was delegated for this specific task.