I think Decode do great work, and I suggested they submit this here. I expect to fund at least part of it, and am chatting details with them.
3
Decode Research - Compute for Generating Dashboards & Autointerp
Not fundedGrant
$0raised
Project summary
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.
What are this project's goals and how will you achieve them?
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.
How will this funding be used?
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.
Who is on your team and what's your track record on similar projects?
Joseph Bloom, Curt Tigges, Johnny Lin of Decode Research. We have generated ~20% of the dashboards already.
What are the most likely causes and outcomes if this project fails? (premortem)
This will likely not fail unless something really weird happens, like all AI compute resources are suddenly 10x the expected price.
What other funding are you or your project getting?
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.
Austin Chen
3 months ago
@NeelNanda I likewise think Decode do great work; as I've mentioned elsewhere, both Johnny and Joseph are two of Manifund's standout grantees and I was excited to see them team up on Neuronpedia and other ventures. Excited to support this grant!
I want to note that the grant writeup here is a bit sparse -- which is okay, given that Johnny and Joseph have a proven track record and are seeking funding for a very specific task; but from a transparency & reporting standpoint I'd appreciate it if the feature dashboards/data for them were published and linked here once this task is done!
Johnny Lin
3 months ago
@Austin Hey Austin, sorry about that and appreciate the feedback. I should have spent more time on this.
I was semi-concerned that there might be details/specifics that GDM might not want us to reveal so I erred on the side of being more conservative, but thinking about this again, the Gemma Scope project is already public so there is no point in keeping details under wraps, and it is definitely better for funders to have more transparency.
I've just added more details and links under the "What are this project's goals?" question and will update afterwards after we've completed this.