First, the headline result of 0.7*sigma improvement is the output of a statistical based on lessons/reviews they engaged with and their mid-term score, with that shift being for "full engagement". Based on their tables something like ~16 students (11% of the group) actually reached that level of engagement
Second, trying to incorporate past grades into their modelling is not a substitute for a randomized trial.
Third, the headline engagement number of 90% is for "engaging with the platform, via Module Review or Lesson Quizzes, at least once". I don't know why much of that couldn't just be attributed to novelty. Or even partly a professor with all sorts of enthusiasm for the platform.
Fourth, the "full dosage" effectiveness is measured based the final exam scores. Were these exam questions produced independently from the "Phosphor" materials? (e.g. by blinding?) Were they checked for direct overlap with those materials? The 0.7 sigma shift is 3 points on a 24 point exam; if even a few of the questions on that exam were very similar to those materials it could account for almost all of it. This is not clear to me from the manuscript.
If this was the case, then it's a question less of "is AI effective" vs. "did the students look at the materials". You could still argue that the AI platform got them to read, but that is a somewhat different statement than the AI helped them learn.
While I echo some of your points, [1] is bad example (as a Canadian).
Research money in Canada is harder to come by; a basic research grant is roughly ~5x-10x lower than a comparable American grant (students are cheaper here, so its not completely proportional, but equipment, travel, etc doesn't scale).
The example for money for poaching international researchers also comes with the asterisk that while they found ~$2B for this, they also are cutting the base funding of the federal granting agencies by a few percent at the same time, atop of that funding being anemic for decades at this point. A big "fuck you" to the Canadian research community in my opinion.
Also a physicist here -- I had the same reaction. Going from (35-38) to (39) doesn't look like much of a leap for a human. They say (35-38) was obtained from the full result by the LLM, but if the authors derived the full expression in (29-32) themselves presumably they could do the special case too? (given it's much simpler). The more I read the post and preprint the less clear it is which parts the LLM did.
Is anyone else having trouble using even some of the basic features? For example, I can open a comment, but it doesn't seem like there is any way to close them (I try clicking the checkmark and nothing happens). You also can't seem to edit the comments once typed.
In my circles the killer features of Overleaf are the collaborative ones (easy sharing, multi-user editing with track changes/comments). Academic writing in my community basically went from emailed draft-new-FINAL-v4.tex files (or a shared folder full of those files) to basically people just dumping things on Overleaf fairly quickly.
Seems like the someone dug something up from the literature on this problem (see top comment on the erdosproblems.com thread)
"On following the references, it seems that the result in fact follows (after applying Rogers' theorem) from a 1936 paper of Davenport and Erdos (!), which proves the second result you mention. ... In the meantime, I am moving this problem to Section 2 on the wiki (though the new proof is still rather different from the literature proof)."
This is a comparison between a new and interactive medium (+ slides, mind-maps, etc) and a static PDF book as a control. How do we know that a non-AI based interactive book wouldn't give similar (modest) increases in performance without any of the personalization AI enables?
This is simply not true, condensed matter physics makes up the largest sub-field of physics (about half by some estimates).
I think the more relevant aspect is that to reach the frontier in a wide array of fields you a solid grounding quantum physics (and several other "new" -- i.e. within the last century or so -- topics) that have displaced more "old-fashioned" topics like continuum mechanics.
I second the recommendation of this book -- it really is quite excellent and covers at an advanced level most of what is commonly left out of a modern physics curriculum (though it is quite imposing in its size/weight).
First, the headline result of 0.7*sigma improvement is the output of a statistical based on lessons/reviews they engaged with and their mid-term score, with that shift being for "full engagement". Based on their tables something like ~16 students (11% of the group) actually reached that level of engagement
Second, trying to incorporate past grades into their modelling is not a substitute for a randomized trial.
Third, the headline engagement number of 90% is for "engaging with the platform, via Module Review or Lesson Quizzes, at least once". I don't know why much of that couldn't just be attributed to novelty. Or even partly a professor with all sorts of enthusiasm for the platform.
Fourth, the "full dosage" effectiveness is measured based the final exam scores. Were these exam questions produced independently from the "Phosphor" materials? (e.g. by blinding?) Were they checked for direct overlap with those materials? The 0.7 sigma shift is 3 points on a 24 point exam; if even a few of the questions on that exam were very similar to those materials it could account for almost all of it. This is not clear to me from the manuscript.
If this was the case, then it's a question less of "is AI effective" vs. "did the students look at the materials". You could still argue that the AI platform got them to read, but that is a somewhat different statement than the AI helped them learn.