Description of subprojects and results, including major changes from the original proposal

I built and launched AI Safety Roster, a searchable directory that helps people find relevant people, organizations, and work across AI safety.

The directory grew from 754 people and 183 organizations to 1,768 people and 650 organization entries. Profiles connect people to their research, projects, organizations, and public funding activity. A public API makes the directory available for reuse.

The biggest improvement was moving beyond browsing profiles to answering practical questions. Visitors can ask who could help organize a community, speak about mechanistic interpretability, or collaborate on AI evaluations. Roster returns relevant people with explanations and links to supporting work. This responds to feedback on the proposal about making the directory useful for concrete decisions.

I expanded profile information using public sources, including Manifund, Longterm Wiki, Alignment Forum, personal websites, GitHub, and arXiv. 1,343 profiles now include linked work excerpts, and 204 distinct papers are linked across 40 profiles. The site also includes 839 AI-related Manifund project listings and 89 selected safety-related Anthropic job listings.

The implementation changed from aiming for comprehensive collection and AI-written profiles toward providing inspectable, source-supported information. Coverage remains incomplete, unsupported fields stay blank, and uncertain identities require review. I also renamed the project from ATLAS following feedback about the existing AI Safety Atlas resource.

I built a repeatable collection and review process. A weekly local refresh gathers updates and potential additions, preserves existing corrections and removals, and holds ambiguous identities for review before publication.

The main lesson was that directory size alone is insufficient. The value comes from connecting people to specific work and explaining why that work is relevant to a visitor’s question.

Thank you to everyone who supported the project and provided feedback. I welcome suggestions for missing people, corrections to profiles, and examples of questions the directory should answer.

Spending breakdown

Development time: building, improving, testing, and deploying Roster - $2,500

AI APIs: profile processing, search embeddings, and question answering - $350

Hosting - $380/2yrs

Scraping and data collection - $120

Claude code - $200

Domain registration and email - $450/4yrs

Total - $4,000

@charbel-raphael Sure thing! Thanks for the note!

@MarcusAbramovitch thanks for the pointer, I dug into Ozzie's work properly. I think we're solving different problems. His own vision doc (https://www.longtermwiki.com/wiki/E883) calls LongtermWiki "a strategic intelligence platform for AI safety prioritization" for funders and researchers asking "where should the next dollar or researcher-hour go", built around cruxes, worldviews and intervention rankings. People and orgs are minor background entities there, and he lists "community features" as a non-goal.

Connecting and matching people is exactly that, and it's the core of ATLAS: the point is the people, and getting them fast to what they need, a hire, a collaborator, a talk, funding. So I see them as complementary rather than the same thing, his data could even feed mine. I've emailed Ozzie to figure out how we can collaborate. Curious where you see the line between the two.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

I've been thinking about where this goes after the directory is solid and wanted to share it here.

The bigger goal is to make ATLAS the matching layer for the field, not just a place to browse:

For hiring: a lab or recruiter posts a role and gets a ranked shortlist of people who actually fit, instead of searching. Same for new labs/startups that just want newcomers or fresh people in the field, we surface them. And the reverse, a person sees the orgs and roles that fit them.

For projects: match projects with collaborators and with funders.

Richer profiles: using the scraper, show real signal on each profile, publications and their actual work and links, so labs can see who they're looking for.

The part I'm most interested in, but it only works with the labs: a portable candidate signal. Right now, every lab runs its own screening and work tests, and every candidate redoes them from scratch. If labs want to co-design or share an assessment that lives here, ATLAS could be the top of the funnel, a candidate does it once, gets feedback on what to work on, and the lab gets a shortlist of aligned people, recently validated. The hard part is labs trusting the signal, so this is more an invitation to labs than something I build alone, and it comes after the directory is solid.

For funding: Surface fundable people and projects to funders based on their real record, to feed existing funders like Manifund.

All of this is stage two, on top of the funded plan. Curious if any of it is useful, and whether any labs or funders would want to be part of it.

@RyanKidd Many thanks!

@Austin thanks for the feedback, it's helpful. I worked on it again: it's a directory now, just a fast searchable and filterable table of real people and orgs. Would mean a lot if you check it again. And thanks for the longtermwiki pointer, it helped.

@ronakrm Done a chunk of this. The default landing is a fast, lightweight table now, with direct search and filters. Feedback button is in :). Semantic search is next on the plan. please check again!

Thanks for the steer ^^.

Hi @Austin, I thought you might be interested to check this out: http://aisa.nahdha.tech -- I welcome feedback!

Thanks!! @aashkapatel