AIs can now pull quotes from articles without hallucination
and maybe transfer knowledge across domains....
Last night I was working on another one of my debate topic briefings and I asked the AI to find some sources for some different potential claims related to a debate on a regional free trade agreement in Latin America.
Not only did it find the sources (expected), but it pulled direct quotes. In the past, it hasn’t done this unless I asked, and when I did, the quotes sometimes (around 60-70%) hallucinated. So, I stopped using that method.
Last night, however, it voluntarily provided the quotations and when I checked on the sources the sources were all of high academic/research quality and none of the quotes were hallucinated.
Academic and policy research and writing is going to change a lot. It seems that someone who knows what they are doing could (co)-write a very good MA thesis in less than a month after the basic study was complete, and now that can potentially be done with synthetic data.
AI isn’t going anywhere. It gets better and better (more “intelligent”) every day.
Pulling quotes from sources is really small potatoes. Today, Google, and its partner, Isomorphic labs, released an AI that can predict the structures interactions of most molecules, including DNA, RNA, proteins, and ligands.
As noted by Jim Fan, we may even have another hint of AGI:
I find this level of generality absolutely mind-boggling. The same transformer+diffusion backbone that generates fancy pixels can also imagine proteins, as long as you convert the data to sequences of floats accordingly.
We are not there yet at a single AGI model, but we have successfully built a menu of general-purpose AI recipes that transfer training techniques, data pipelines, and neural architectures *across domains* that don't seem related.
Today’s freshman will graduate into a completely different world than the one they entered in their freshman year.
