They say that pressure makes diamonds. But slowly depositing carbon atoms on a seed crystal also makes diamonds.

it is remarkable how the majority of these awardees are in venues with minimal viewership/readership.

Also there's a lot of conflation here on Pangram: 1. Writing with the help AI and having AI write it 100% for you is not the same folks. 2. As a general reminder, if you write youself and ask AI to fix spelling, Pangram WON'T flag it. 3. If your writing is 100% AI on Pangram, it means you copy pasted and are intellectually lazy. This is why the cope of "Oh I just write my captions with AI" doesn't work because you did not write it!

1. Many LORs were slop even before AI 2. The amount of mental effort you need to write a comprehensive LOR, if you actually know the candidate, is not astronomical, especially if you have minimal experience writing these letters. 3. One should never disclose online that they are putting confidential communications through pangram

In academia, the early bird generally does not catch the worm, when it comes to submitting something with a deadline. Submitting close to the deadline (with sufficient buffer to allow for portal problems) has no disadvantages, and sometimes has advantages (e.g. being able to include products/awards that materialize just before the deadline).

You would think that the emails sent to professors to submit LORs for grad school have the application deadline on them, and you would be wrong.

At freshman orientation 2002, Stanford touted Carly Fiorina (medieval history major) as their pinnacle inspiring alumna (this was before she ran HP into the ground)

It is important to maintain a healthy fear of being roped into a new committee when deciding whether to micromanage colleagues.

- Internal model X, produces slop too expensive to verify - Dump slop over to the maths community, let them critique & verify the work for free - Feedback into RL loop for next "internal" model Repeat until the maths community is exhausted & no one wants to do maths anymore.

One thing that seems lost in this chaos is that mathematicians’ use of LLMs has been consistently productive over the past year or so, leading to carefully written papers that (1) contain results of a comparable level to those produced in the OAI drop; (2) properly acknowledge both AI use and other people’s work; and (3) are ready for peer review and dissemination at workshops and seminars. And all this was done using models that, apparently, are less capable than the ones used by OAI (there is no way to verify this independently, hence the “apparently”). So, what is getting lost in the chaos is that there was already a way to go about this intelligently, without damaging the human infrastructure that produces mathematicians, and with mathematical progress picking up speed. A clean way forward would be something like giving mathematical researchers access to frontier models. This would give OAI independent, external validation of the quality of their models, while allowing mathematicians to produce papers that are readable, credible, correct, and fair to both the literature and the developers.

People with deep expertise in a specific niche are very sensitive to the uncanny valley when AI generated text is not obvious slop but also does not make a strong scientific case (while cosmetically looking like it says something specific). People in an adjacent field consistently read these AI-generated documents more favorably. It’s fascinating seeing this play out in panels.

Sensitivity to AI slop is sort of like how when you have deep expertise on a topic and you hear someone you thought was smart opining on it, and you realize they are not that smart. Scientific writing is a skill that is acquired during training, crucial for organizing and justifying your supposedly novel ideas. Some people could read a few dozen research papers in their field, feel the vibe, and write something in a similar style instinctively. Others got to the same place with many iterations of practice and feedback, or with some formal training. When you have that sensitivity to how scientific writing is supposed to read, AI generated text is very jarring and exhausting to read through. There are a lot of copes about Pangram, but if you use AI to clean up your grammar, shorten the text, or translate from your native language to English, it will reflect minimum AI content, and will also not be jarring to an expert human reader. It is not fair to demand human time to evaluate something you did not write yourself. If the output is a proposal, how can you execute the research if you can’t articulate the research? When search engines came along, it did not obviate the need to memorize; it only made it clear that if you don’t know anything, you don’t know anything, even if you can theoretically look it up (but how do you know what to look up). Similarly, the possibility of cosmetically outsourcing your thinking does not diminish the value of having your own thoughts (especially if your job is having novel ideas).

My department is doing a search in experimental condensed matter! recruit.ucdavis.edu/JPF07916

The bioethics circle jerks during the Human Genome Project, which were allocated 5% of the project's funding, were a total waste of time. They had countless conferences and workshops imagining bad things that could happen like punitive health insurance rates based on genetic risk, none of which actually came to pass. Meanwhile all the real issues that arose from ubiquitous DNA data, like accidentally doxing your criminal relatives with your own 23andMe data, were never imagined. It's a condemnation of the precautionary principle. Philosophers can make up all sorts of hypothetical scenarios in a vacuum but it's nothing like actually trying things and seeing all the weird second and third-order consequences that emerge and *then* dealing with the unwelcome ones.

Important concepts that I learned only in materials science courses, not solid state physics: point groups and space groups, all the tea on dislocations, mechanical properties of solids beyond one lecture, actual curriculum related to liquid crystals beyond appropriating the terminology.

We're hearing the same story from universities in both Canada and the US. Students are arriving unable to do basic math. But the problem doesn't begin in university. People tend to underestimate just how hierarchical math is. Learning new math concepts depends on what came before. When students don't develop a strong foundation early on, the gaps don't repair themselves. The gaps just get worse. Then students are moved ahead when they didn't master the foundational material and later math is inaccessible. It's nearly impossible to learn more math when your foundation is made of sand. University is not the best place to try to fix these issues. We need to look further back and address the core issue. And I believe in many cases kids are being done a disservice in especially K-8 math. I think this is one of the most serious issues facing our society right now.

Deanlets are the ultimate proof of Pournelle’s Iron Law of Bureaucracy. Faculty hiring stalls and positions are eliminated while associate deans, directors of belonging, assessment coordinators, etc multiply. Success becomes administrative continuity over student learning.

A path-dependent journey

I tried to sign up for the Claude Team plan for Scientists last week, and they could not verify me as a PI based on my website and google scholar (so much for AGI). I tried again with my website, my department website, my orcid, and my NSF grant--approved! Was it the extra info, or the bad publicity from the failed psyop that happened between the two attempts?

To add a bit of clarity about who proved what. Tristan + Levent: 3d incompressible Euler with forcing OpenAI: 3d incompressible Euler without forcing OpenAI: Navier-Stokes with forcing No one: Navier-Stokes without forcing Euler equations = Navier-Stokes without viscosity. Forcing means external force. Absence of viscosity and presence of external force make blowup easier to construct. So in the 2x2 grid Tristan+Levent ticked the weakest box, OpenAI ticked the two next weakest, and the strongest is still open. Navier-Stokes is not over.