Melanie Mitchell

💯 Same comment for physics professors wondering if they need postdocs any more. (I have heard this comment from more than one.)

@GregKamradt I use em-dashes and colons all the time, sigh. AI is stealing our best punctuation marks. Who can we sue?

My current smells of AI slop/writing: 1. Use of em dash "—". I haven't seen anyone seriously use this over a hyphen "-". Double points for wrapping — or making double points in a single sentence — give it away 2. Making a statement and then colon: like this 3. More subtle, but tone that is off-character for the authors

Exciting new initiative for @dileeplearning and @AsteraInstitute to dive deep into neuro-inspired AI:

@TomChivers Not sure what Pollan means here, but maybe he's referring to this kind of computational complexity result? quantamagazine.org/how-computatio…

just saw this in a review of Michael Pollan's new book theatlantic.com/books/2026/02/… surely even a second's thought would show this is wrong? A single neuron cannot, for instance, tell me what move to make in chess or the best route from Crouch End to Muswell Hill

@GregKamradt @mikeknoop @ZennaTavares Indeed, this is the right direction. Understanding is not always reflected in accuracy.

Thanks for this @MelMitchell1 I believe what you’re referring to is along the lines of @ZennaTavares has done with AutumnBench? Rather than focusing on a score, have an AI system interact with an environment and then ask it questions about that environment to ensure it actually understood it, not just stumbled its way to a high score

I appreciate this reflection on why ARC matters, but in the replies I expressed some reservations about what high accuracy on ARC benchmarks actually reflects for assessing "human-like fluid intelligence" (Chollet's stated goal for ARC).

@mikeknoop I wrote about this more comprehensively at aiguide.substack.com/p/on-evaluatin… Thanks for all your work on supporting ARC, and I would be happy to discuss any of this with you in more detail.

@mikeknoop To make sure that progress on ARC (including 1, 2, and 3) correlates with actual progress in human-like fluid intelligence, rather than reflection of "Rube-Goldberg-like" shortcuts, evaluation has to be broadened to understand mechanisms underlying high accuracy....

I wrote about the late philosopher Brian Cantwell Smith, and his profound thinking about AI and the nature of intelligence. aiguide.substack.com/p/on-brian-can…

@alfairhall Bizarre. But also... as far as I can tell, 100% male.

Epstein list of invitees to 92nd St Y. So statistically unlikely that so many ……. have blue eyes

@tdietterich @arxiv

@tdietterich @arxiv Everyone: You can use ChatGPT for this! Here's GPT5.2 on four independent trials:

#ICML2026 authors, if you can flatten the curve on submissions to @arxiv , the cs.LG moderators would really appreciate it. Today we have 4 times the usually number of submissions. Please select a random number unif(1,14) and wait that number of days to submit! :-)