Update on LLM reviewer situation:

PM is down to let us pitch them our argument. Good news: PM seems like a cool person, is open minded, and is being pretty frank about the forces at work here. Bad news: taking action on this will open a whole can of worms, so any proof has to be ironclad. After conferring with our local grant wizards, the battle plan is to crank out a 15 minute pitch consisting of:

• a 2 min elevator pitch of our tech, highlighting what the reviews mangled
• intro to LLMs for people who know what glycosylation is
• intro to semiotics for the same
• show how transformer architectures transform symbols into symbols to produce text-shaped objects without actual intent, ideas, or context (and why “automated AI detection” is also bullshit).
• show a few examples of plausible-at-first-glance gen-ai slop (the nonexistant turkish fortress, mouse dck, etc)
• Highlight how our weird reviews (both good and bad) fit exactly into this bin (absolutely mis-interpreting a table, inventing a bacterial species we didn’t use and talking shit about it, miscounting our team members, etc)

We’ll be leaning on the Stochastic Parrot paper pretty hard, because it’s a good entry into the field on the skeptical side and is just well constructed in general. I’m also on the hunt simplified diagram for how LLMs convert tokens to arrays to tokens from the original transformer literature. Unfortunately, so much of the literature is obscurantist on purpose, and I want to avoid falling into the “It can’t be that stupid” trap. Any pointers in that direction are most welcome!

Wish us luck, heh!