Thanks Johan - makes sense. Congratulations on this work.

This is super cool Johan. What is your feeling on the reason that this works so well?

arxiv.org/html/2412.07...

According to this, at CVPR 2023, 1% of Authors account for 50% of the papers… is it really skewed this much?

Results get way better! What’s going on? I look at code added and LLM has added class labels + random noise to the vectors fed to the classifier. Adding the class labels would be a mistake I could accept… but adding some noise to them - that’s a bit suspicious! 😬

Interesting experience. Done with experiments and ask an LLM to look for any data leakage (surprised by results). LLM suggests to add some checks and balances so I defer to its grand intelligence and allow some code to be added. 1/2

Ok - I guess I totally missed the point. 😂

At least I learned something cool.

🤔… Kersten?

Peyman: I must say that your historical perspectives and observations are some of the most interesting and insightful things online. It strikes me that important context is lost to time, but retains a role in supporting new discoveries. If you had a book full of these, I’d be first in line to buy it

Successively melted and refrozen snow against a building with corrugated metal exterior?

Manus, make me some slides about generative AI in the style of @csprofkgd.bsky.social

In case you’re wondering where many of the globally important AI conferences in 2025 are being held, here’s a quick reference:

🇺🇸 NeurIPS ‘25
🇺🇸 CVPR ‘25
🇺🇸 ICCV ‘25
🇺🇸 ICRA ‘25
🇺🇸 AAAI ‘25
🇺🇸 NAACL ‘25
🇺🇸 WACV ‘25
🇸🇬 ICLR ‘25
🇨🇦 IJCAI ‘25
🇦🇹 ACL ‘25
🇨🇳 EMNLP ‘25
🇹🇭 AISTATS ‘25

Is aideadlin.es no longer updated? Does someone have an alternative I can bookmark?

Rebuttal:

I agree that the system is *very* noisy now. But it’s always felt like a bit of a lottery. The paper was rejected with a 4th reviewer giving it a 4. It’s published as a journal paper instead. I sympathize with everyone holding a lottery ticket… but keep buying your tickets! It’s worth it!

10 and 1 had the qualifier that if the paper were (accepted/rejected) the reviewer would consider not reviewing again. The 1 began with “This entire manuscript consists of pretentious rhetoric”. Reviews are certainly noisy now… but it’s always been a bit of a lottery. (2/3)

This post is in response to some of the discussion I’ve seen around recent ICLR reviews. I wish I had an archive of my emails from previous institutions to have all details. About 15 years ago I had NeurIPS reviews with scores 10, 5, and 1. Regarding 10 and 1, assigning these scores… (1/2)

It feels like computer vision is evolving in a direction where more and more of the most central advances rely on advanced mathematics. Is this my imagination, or do others feel this as well?

I had to give it the full exaggerated Canadian stereotype flavour. Now we’ve got a moose, a beaver, maple leaf adornments (and possibly fleur-de-lis?). Also 3 of the pieces look to be bottles of maple syrup including the beaver. 😂

Hello Bluesky!

Congrats Kosta! 🎓

Bluesky user counter web site:
(In case you’re interested in growth rate)

bcounter.nat.vg

Starter packs help a lot. Would love to see the limit on size of starter packs being raised. Assuming there is mostly a direct mapping between BlueSky names and the other place, it would be cool if there was a way to rebuild one’s own “followed” through a personalized starter pack.

Awesome Kosta - thanks! Based on how your list 2 is growing, maybe you’ll need a list 3 soon. 😂

I thought this might be coming. Given your standing in the community, I think it might be worth an AWESOME list 2. And if anyone reading this hasn’t followed Kosta yet - do it! You won’t regret it!

In the case of transformers (and especially ViT), it seems like interesting possibilities abound in getting creative with different choices for Q, K and V. Flamingo is one nice example of this. Any suggestions on innovative works that do something outside typical “vanilla” configurations?

If you build it they will come (given the right reason). Then the money needs to flow. Good test though - would love to see the same in 3 months.

Thanks so much Kosta - you have no idea how much I was hoping for a list like this. It was the missing seed in starting to regrow the network.

Agreed. This is one of my favourite books. Irrespective of how many of the specific details are exactly right, the overall point is compelling.

If this traction continues (this platform and your engagement and solid content) surely you’ll get there. I’ll be there for the party to celebrate 100k.