I feel seen

Au contraire. C'est une nouvelle ère qui s'ouvre pour les mathématiques. Une ère où la démonstration formelle est largement automatisée et où l'accent sera reporté sur le développement de nouveaux concepts, nouvelles abstractions, nouvelles définitions, et nouvelles conjectures. L'invention du bateau a réduit l'importance de la nage, mais a permis la découverte de nouvelles terres.

OpenAI seems to have made a serious breakthrough toward the Riemann hypothesis, and people are still virtue-signaling their “pure human” work with Pangram scores.

Random thought: One unexpected benefit of replacing university administrators with agentic workflows: we may finally find out what the policies are. “Discretion” sounds noble until you realize it often means that the written rule is optional, the real rule lives in someone’s head, and the outcome depends on who you are, who you know, and whether the administrator likes you. Favoritism with a respectable name. An agent cannot run on vibes, relationships, institutional status, and “we handle these case by case.” Write the damn policy down.

Give a little bit of power to people and see how they abuse it. Seriously, write a good review and desk reject with reasons. If you want to reject based on gut feelings and quick scans, just decline the job because you are not fit for it.

No need for meat proxies. AI can already produce reviews that are as good as, or better than, most human reviews. Adding a human in the loop just adds three months of latency without improving quality. Let’s stop pretending otherwise and move on.

In my class I gave students an oral exam, where they had to discuss with a voice AI agent their (already graded) submission. To keep the assignment low pressure and low key, I told them that it was pass/fail, and the only requirement to pass was to take the exam. It did not matter how well they could answer the questions. They could take the exam anywhere, at any time, using their computer. Even in this setting, where there was NOTHING at stake, with the assignment being optional, a few students decided to “cheat.” I could hear them in the recording of the exam being silent for 20–30 seconds after each question, hear keyboard typing sounds in the background, and then a perfectly fluent answer (reading the answer from the LLM output, I presume). I kind of felt depressed afterwards. I kind of get cheaters when the incentives are aligned to push them to cheat. But in this case they had zero incentive to cheat and zero reward for doing so. No, I did not prohibit them from using an LLM to generate their answers for the oral exam. No, I did not pursue cheating charges, and I do not even know whether this is technically a violation of the honor code, given the setting. But I did not think that I would have to put such terms or restrictions in place given the nature of the exam. This is almost like giving your phone to a friend and asking them to record a run on Strava. Why would anyone do that?

On AI grading: It is becoming abundantly clear that AI grading is better than human grading. But at the same time, assessments done with unrestricted access to AI are fundamentally vulnerable to student cheating. AI grading and tutoring are fantastic for formative assessments, where the goal is to help students learn. For summative assessments, where we assign a grade based on how well a student knows the material, we will have to go back to high-stakes exams, heavily proctored, with controlled (or no) access to technology. Universities will develop parts that emulate testing centers. My worry is that the alternative is the path of least resistance. Let students use AI for everything. Let AI grade everything. Everyone gets good grades. Students are happy. Faculty are happy. Administrators are happy. And universities gradually become clubs: places where students socialize, build networks, and establish lifelong relationships, but where nobody seriously measures how much they actually learned. Perhaps this is the current status as well and I am in denial.

First time I’ve reached this level of confidence in AI grading: The TA had already graded the assignment and had done a good job. Not unexpectedly feedback was scarce and limited to why points were deducted. I still asked OpenAI “dot” to regrade everything, find anything the TA had missed, verify the AI-generated answers using Claude, and then revise all grades and feedback directly in Brightspace. I don’t think grading will be among the tasks of any of my TAs moving forward. Assuming we still get to have human TAs in the future.

I’m increasingly convinced that academics on social media should disclose financial conflicts of interest. It’s hard to take a passionate argument at face value when you later discover that the person owns equity in a company that benefits from that position, or is being paid as a consultant or expert witness. If someone is running a startup, the startup is often the result of that passion—and the connection is usually pretty obvious. But when the passion is driven by a hidden financial interest, failing to disclose it starts to look like malpractice.

157,240 workers. 167 countries. They told us who they were, one HIT at a time. As Amazon Mechanical Turk closes, a final tribute from my Mechanical Turk Tracker and a thank-you to the people behind the data. 410 Gone.

Sep 30, 2026 · 20:00 UTC Last HIT: 3N7PQ0KLI4O69Y3NFCLSOSD5LIW3EN Last worker: A25PNZS508M860 API status: 410 Gone As Amazon MTurk closes, my Mechanical Turk Tracker signs off after 17 years. Rest in peace, Mechanical Turk.

We're hiring Applied AI faculty at UChicago Booth! This is a great fit for AI researchers who want to go into academia but don't see themselves being traditional CS professors (e.g., managing a big lab of students). apply.interfolio.com/194344 You might be right for this if: (1) You want to take the IC track. Industry has realized that talented researchers don't all want to be managers, so most companies have an Individual Contributor (IC) track that you can stay on as you become senior. This typically isn't an option in CS academia---your job from PhD -> faculty often becomes very different, one of funding and managing a lab. Lots of people like this, but a lot don't. At Booth, it's possible to stay an IC researcher---or run a much smaller research group if you want. (2) You want to not worry about funding. Almost all of my funding comes directly from the department, so I don't have to write many grants (in fact, you could choose to live an entirely grant-less lifestyle if you wish). Speaking to my PhD friends who went on to top CS departments, this is a major source of pain in the current funding environment. (3) You want to do slow science. We emphasize quality, not quantity. We want people who might only write one paper a year, but whose work goes on to have a big impact on the field. Impact can be measured in many different ways as well, not just by the number of citations you get. I'm entering my second year at Booth and have really enjoyed my time here, as someone who was in a CS dept from undergrad (UToronto) through PhD (Stanford).

Go back far enough, and every ancient person is either: a. ancestor of everyone alive today b. ancestor of nobody alive today If their lineage survived, they are your ancestor. And mine. Far enough back, we all have the same ancestors, just mixed through different paths.

LLMs fundamentally solved the “common sense” and “everyday context” problem in AI for most knowledge tasks. We are still digesting how much that makes possible. The “stochastic parrot” crowd was wrong, plain and simple. The revisionist attempts are laughable. Bringing common-sense knowledge to all sorts of tasks is precisely what makes LLMs useful. Take that capability away and see how much still works. Common sense in the physical world remains unsolved. That is the next major research frontier.

Just curious: has anyone run a paper through a non-Gemini AI reviewer and gotten an “accept” or even “minor revisions”? I’m trying to understand the panic about “AI slop” flooding journals and conferences, when AI reviewers can filter out technically incorrect work. They find plenty of flaws even in papers people have spent months working on, including mine. So I have some trouble believing that a mostly automatically generated paper will just sail through. Perhaps my work is below slop, idk.

The Pangram idiots cannot read the text to decide whether it is worthwhile or not but worship the Pangram score.

On the French AI novel: so many reflexively claiming bias or saying "you can't detect AI writing". I will donate $1000 to the charity of your choice if you show me a single literary French text of 1000+ words, pre-2022, with Internet Archive link, that is 100% AI on Pangram. 1/2

Looks like we have over 500 submissions to ICLR this year. openreview.net/group?id=ICLR…