Hahaha, I apologize for the exaggeration here, @NeelNanda.
Here's my logic: there are numerous stories about Indian students discovering cybersecurity during their school time through a workshop or session and then pursuing that path, either along with schoolwork or as soon as they enter college. I want to create a similar kinda impact with AI safety. I know after the workshops or the DPLS (Daily Peer Learning Sessions), students won't immediately become a Neel Nanda and start working on mechanistic interpretability, but that early discovery of future Neel Nandas (the most appropriate Indian personality that I could think of while drafting this was yours 😅) and guidance in the right direction is super important. And the cost mentioned is for the discovery of those Neel Nandas.
Here's what I've seen that makes me think safety literacy changes the trajectory even for AI builders and AI capabilities researchers: I was at the Claude 2 hackathon in San Francisco and spoke with Emmanuel, who was an AI capabilities researcher at Anthropic at that time. I asked why safety was handled through fine-tuning rather than baking it into pretraining. His answer was because doing it during pretraining would cost you the capability. He wasn't dismissing safety; he just didn't see another path at that time. The thing is, someone who deeply understands safety has a completely different map of what "another path" might look like. They're more likely to notice a promising solution that balances both (profit and safety), and more likely to pursue it, than someone who has never thought about safety at all.
With Phase 0, there's not much quantitative data, but there is qualitative data (testimonials and the interest quotient). Will try to back this whole thing with Phase 1 quantitative data soon and present it on the project. Also, would love to have you as a speaker in one of our sessions to guide and inspire our students :)
Thanks so much for your valuable attention; it really means a lot ❤️
P.S.: The project budget and numbers have been updated multiple times since the initial draft after multiple conversations with funders, so they reflect a ballpark figure and not an exact figure