grantmaking.ai Launch Round
This is a one person project, I have already built a rough prototype and a detailed vision doc .
This project aims to help the new talent field builders is attracting, in reading and understanding research papers in the field in the most frictionless way.
The project aims to do this by creating an interactive and traversable written version of the papers of the field, which can help them understand the same papers at whatever level of depth or complexity regardless of their existing knowledge about the field. It is enabled by the fact that this tool makes the prerequisites and concepts easily navigable unlike a static paper.
The reason this project can be a highly leveraged one is that as the models become more capable, technical, engineering and implementation is going to be commoditized whereas insights and overall understanding of the field are going to be highly valuable.
Chris Olah and others have identified a problem of lack of distillation of research output. They regard interpreting new research as important as novelty, Either this has to be done by the author by producing a write-up about the research or by thousands of readers separately at their own expense. His attempt at distill. pub was addressing this issue. However, this did not scale as there are not many capable researchers that are available to do this work.
Llms also fail at distillation as they are more prone to producing explanations that 'looks right' , that actually 'is right', especially at a paper scale . The readers who want to understand a new paper usually upload a paper to an llm and endlessly prompt it to understand it , but it has two issues ,one , llms' context windows are limited. So it cannot have 1000 papers in it, even if it could , the needle in the haystack problem make this difficult to make connections and find concepts at the right granularity. Also, the prerequisite problem is brutal for new entrants into the field as raw llm prompting or claude code agents may not necessarily track all the concepts correctly, even if it could, it is going to take considerable effort from the researcher just to understand the prerequisites
What we are planning to do is to break down the papers to its granular concepts , relationships and connect it to existing corpus using llms , at the current level they are good at handling granular and atomic concepts. that way their effort in reading a paper is significantly reduced, and the prerequisites can be easily traversed in an interactive way which a paper is not.
This will accelerate the huge talent pipeline that already exists in AI safety. New researchers will be able to produce substantial research on a much faster schedule, strengthening the field and its research output as a whole. The case for this grows more important as models get smarter. To leverage their technical and implementation skill, and to keep pace with capabilities, the field's process of sense-making has to be streamlined. This project aims to contribute directly to that.
The minimum version covers API credits, crawling the papers, and supporting myself, with a few live canonical spines to start.
The maximum version evolves into the preeminent tool and research explorer for the field, with the canonical spine updated live and competing canonical corpora maintained side by side, so that any bias introduced by the selection of the canon is counteracted.