Open research culture is the best thing to have happened to science. Because of open research, there’s more to read. And, subsequently, not enough time to consume it.
Today, I asked twitter what they would find most useful for managing the research firehose. I used a creative trick - imagine if you had infinite humans working for you. This trick is useful, because it prompts people to act as if AGI is here. (it is)
I’m sharing some thoughts on some particular ones, but all of the ideas were great!
Have the infinite humans create a 2 minute video with visualizations
- choomba
This can already be done with audio summaries. But, one could imagine using the multimodal systems that we have today to make it visual.
Elsewhere in the thread, kosenjuu mentions that a platform that encourages researchers to present things in bite sized form. Research papers aren’t designed for easy absorption - they’re designed for maximum verifiability.
AI systems can help filter down, and change what function the information serves.
Have an agent sit on top of the research stream, and compile a list of summaries that a user can read each morning
- AlignmentGuts
The information is unmanageable. We’re dealing with it through the collective meat based filter of internet friends, yielding signal when necessary. But, this is incomplete, and things that I’m interested in might not overlap with what is shared.
An automated compilation can be made contextual. I only care about certain types of papers. Even within the papers, I only care about results that have utility for my purposes.
Could we port the filter from meat substrate to silicon substrate?
Customized research timelines - being able to generate a genealogy of a particular subject
- madmaxbr5
When I'm trying to build a prototype for a client, I need to figure out the trajectory of the performance of a particular method, and how the performance has changed (and is projected to change). A timeline would be incredibly useful.
This isn’t just CS research. Pharmaceutical companies spend a lot of resources having researchers carry out surveys on particular drugs. There’s a wealth of information in papers, forgotten in the pile. Is there a cure for cancer sitting in some paper written in the 90s?
horizontal intelligence
The cost of horizontal intelligence is crashing. A single agent may not have much depth, and can’t replace a single human. However, humanity's capacity for repetitive, toilsome intelligence work has exploded.
What kind of work are we going to map reduce?
Raw notes at yacine.ca
Are my notes/blog posts useful to you? Throw me a one time money upvote!
I think you touched on an important problem, in which it's value is likely to increase in time. Many people define the "singularity" as a moment in time where we can no longer keep up with innovations and technological advancement. (I recall this definition being mentioned in "The Gentle Seduction"). Maybe multi-modal systems can help. :)
Individual preferences in summarization and its implication have been a blessing to the human race (re learning to summarize from HF + recursively summarizing books, etc). Makes me wonder if we will have something close to "infinite humans" when we are willing to provide more signal and let models cook.