Quote: "The Overhang consists of the unrealized capital gains of past mathematical creativity, the latent value from connecting the dots in the existing corpus. It is a dividend of canonization. Mathematician X states problem A, mathematician Y crafts concept B, then mathematician Z notices that B trivially solves A and “captures” the social reward. But in the process of capturing the reward, Z usually introduces new concepts and new open problems, reinjecting latent value into the Overhang.

LLMs can be trained on the entirety of the mathematical corpus. Thanks to their phenomenal memorization and pattern-matching abilities (without always being able to map out their associative logic and attribute due credits), they are in a unique position to harvest the Overhang. By contrast, professional mathematicians have typically read a few hundred articles in their career, out of millions of existing references, less than 0.1% of the total.

This will lead to great discoveries, which is unambiguously exciting. But it could also lead to a sad new deal, where human slaves painfully curate the Overhang while AIs systematically beat them at the finish line."

source: https://substack.com/inbox/post/183753276

"The negative learning effects are larger for students with higher initial achievement. The differences in the estimated full (6-10 month average) effects are substantial, with a 50% gap between the most negative effect (-24 percent) for the highest tercile and the least negative (-16 percent) for the lowest tercile. "

Not top 10% as you asked, but the closest to what you asked. My working hypothesis is that top performance is highly correlated with willingness to work hard, and AI decreases the motivation to work hard.

"At the other end of the distribution, AI students who spend more than 65 minutes on their homework receive homework and exam scores similar to those of non-AI students, suggesting that these students do not use generative AI for homework assignments. However, this group consists entirely of students who adopted generative AI no more than Öve months. Six months after adoption, no AI student spends more than 65 minutes completing their homework (see Figure A5). This is consistent with the gradual process of learning how to use AI tools. It also suggests that AI crowds out the highest level of e§ort."

"Interestingly, in the range of 50-65 minutes, the median and the interquartile range of exam scores of AI and non-AI students are similar. This implies that, in the range where AI students and non-AI students have overlapping homework times, students who spend the same amount of time completing homework on average receive similar exam scores."

"This pattern shows that students who spend the same amount of time on homework learn similarly, with or without generative AI. In other words, generative AI reduces time spent learning for the majority of AI students but not learning efficiency for those who spend the same time studying as the non-AI students."

https://youtu.be/LDSwP37y_W4?si=fMIsdQ2yjoiGChb6

The idea that if only all professors stood their ground then somehow students will be motivated to study doesn't pan out in practice, though. There is already a significant number of students who are perpetually struggling. They are missing basic prerequisites, and instead of catching up on them, they repeated try and fail at learning the same materials, passing only when they got a lenient instructor. The problem compounds because failing brings helplessness and exacerbates their mental issues, which brings more failing. The university cannot sit on their high ground and watch these students struggle, especially if their number reaches a critical mass.

I think you inferring my motivation for the rant and creating a strawman yourself.

I do agree that directing my rant at the generic "fans" is not productive. The article Tao wrote was a good example of communicating the result. I should direct my criticism at specific instances of bad communication, but not the general "fans".

This is a really good example of how to use the current capabilities of LLM to help research. The gist is that they turned math problems into problems for coding agents. This uses the current capabilities of LLM very well and should find more uses in other fields. I suspect the Alpha evolve system probably also has improvements over existing agents as well. AI is making steady and impressive process every year. But it's not helpful for either the proponents or the skeptics to exaggerate their capabilities.