I have no idea about the grant process, but could they implement system in which allowing reputable scientists and researchers vote on grants in their area of expertise?
Most grant awarding processes operate similarly to bids for tender. That is, the awarding body invites researchers to submit proposals that fall within its remit.
The submitted proposals can involve including a year or more of "pilot data", being experiments showing that the proposed approach is feasible. In addition 3 or 4 months of writing and admin (budgets,legal requirements, etc) are needed to complete the application.
So an application can easily be 1.5 to 2 man-years of work to prepare.
A screening process then takes place. This involves the awarding organization vetting the credentials of the applicants, as well as recruiting specialists to give their expert opinion. This takes at least 6 months.
Then there will be panel of experts who review all the above and vote. The vote is "taken into consideration" by the executive of the awarding body who make the final decisions. They might award funding for three years of work for maybe 10-15% of applicants.
As you can see, it is a massively burdensome process, with typically a very poor return on investment on the behalf of researchers.
And no, I don't think current AI is anywhere near competent enough to replace most of this process, apart from maybe the admin and legal sides.
Massively burdensome?
No, it is an absolutely essential process that over the last 80 years of “industrial academic research” has brought us wonders and insights.
Three to four months to design. write, and refine a 13 page document is working at a very leisurely pace in my opinion (43 years of writing more than 100 applications).
And yes, I do think AI systems can improve much of this process and result in much stronger science.
You forgot to include the 1-1.5 years of pilot data needed for a typical biomedical science (wet lab) grant proposal.
Essentially, the overall return on research investment for academic biomedical researchers themselves as a whole is negative. Which is why a substantial subsidy in the form of underpaid students, etc, is needed.
There are a few top performers who achieve a positive return from grant applications. A more streamlined and less fickle review process would be indeed improve matters, though I very much doubt LLMs in their current state of development will help this process.
This is sort of how it works, but usually there is a process of reviewing and asking questions during interviews rather than just voting. In any case, both voting and reviewing require one to read a lot of grant proposals. The problem is that an overworked scientist reviewing grants may just resort to LLMs for that.
This is disallowed by NIH for confidentiality concerns. But reviewer are free to submit the applicants paper to an LLM for commentary. Useful to detect traces of p-hacking that require careful teading.
Most grants in the U.S. are discussed and scored by reputable scientists and researchers in a committee format, and those scores weigh heavily on the likelihood of a grant being funded.
It's actually one of my favorite parts of the job, getting to read new, promising ideas that I think might genuinely help humanity.
The submitted proposals can involve including a year or more of "pilot data", being experiments showing that the proposed approach is feasible. In addition 3 or 4 months of writing and admin (budgets,legal requirements, etc) are needed to complete the application.
So an application can easily be 1.5 to 2 man-years of work to prepare.
A screening process then takes place. This involves the awarding organization vetting the credentials of the applicants, as well as recruiting specialists to give their expert opinion. This takes at least 6 months.
Then there will be panel of experts who review all the above and vote. The vote is "taken into consideration" by the executive of the awarding body who make the final decisions. They might award funding for three years of work for maybe 10-15% of applicants.
As you can see, it is a massively burdensome process, with typically a very poor return on investment on the behalf of researchers.
And no, I don't think current AI is anywhere near competent enough to replace most of this process, apart from maybe the admin and legal sides.