Theo Sanderson

@theo.io

researching viral pathogen genomics at @lshtm.bsky.social, building @pathoplexus.org | theo.io / sandersonlab.org

Had a UKRI funded grant proposal rejected at the "AI triage stage". "This AI-based triage was used only to identify approximately 50% of the strongest proposals to take forward to full human review." 🧪 #academicsky

I am writing to let you know that, unfortunately, your application was not successful.

Review process

As part of the initial triage of the 179 proposals submitted to this call, each was assessed using an AI-based review. Every proposal was read independently by several different AI models against the same seven criteria used to shape this call (including cyber security relevance, research quality and novelty, importance of the problem addressed, feasibility of the project plan, likely outputs and impact, value for money, and responsible research practice), with each model asked to give a score and a written justification with supporting quotations from the text. To guard against any one model's idiosyncrasies, we also ran a second, independent process in which different AI models debated the merits and weaknesses of each proposal before reaching a judgement, again against the same criteria. Scores from both processes were combined to produce an overall ranking. 

This AI-based triage was used only to identify approximately 50% of the strongest proposals to take forward to full human review.  Projects selected for funding were drawn from those that progressed to the full human review stage. We recognise that AI-assisted assessment is a new and evolving part of the review process, and we are continuing to evaluate how well it aligns with the judgements of human reviewers on this call. We will soon produce a document detailing our method, so that others can build upon it and improve it.

Feedback on your application

Your application was not selected to progress beyond the AI triage stage.  To provide transparency on the outcome of this assessment, we are sharing below the scores (out of 5) from each of the two AI review processes described above, together with the mean score across the two processes.

Anthropic do geniunely think AI may kill us all. They believe that no one else is going to even try to make a reliably benign AI, so they have to be first. It definitely sounds sort of insane, and I’m nowhere near equipped to tell if it is. But the belief is sincere and internally consistent.

I wanted to submit a grant proposal this week and ended up not doing it precisely because of this. I have no idea how in a project that, say, lasts from 2027 to 2032, I can promise results that cannot be achieved by an LLM in 2031 with one person prompting.

Ragnar {Groot Koerkamp}@curiouscoding.nl · 3w ago

Writing TCS grant proposals is gonna be fun when you can just give the thing to an LLM and get everything done in a few days, and at the same time you have to convince the committee that no, this is not to much work for 1 person for 4 years.

Who knows if relevant here but it's intriguing to be moving towards a world where all customers of each AI firm are co-opted into a hive mind

Bild

Astra implemented a feature I'd been wanting in gensplore.genomium.org for a while but hadn't managed with previous LLMs - overlaying a sequence onto its reference. The button isn't very obvious, so as an experiment I also asked it to make an (admittedly annoying and slop-py) video walkthrough