this is now how i now imagine bloomington in my mind

i bet somebody already distilled this into a skill and fed it into their proof agent

what did sutton see?

The funniest part is that AI has apparently learned both how to write a paper and how to reject one.

But neither necessarily requires understanding the science.

Generating plausible research and generating plausible criticism are both much easier than discovering something genuinely new or evaluating whether it matters.

We're automating both sides of peer review without solving the hardest part: scientific judgment.

Maybe we've achieved recursive self-criticism before recursive self-improvement.

so you’re telling me that, simultaneously:
(1) most ICLR submissions are ai generated
(2) most ICLR reviews are ai written
(3) the median ICLR score is a 2 (reject)

so either ai has learned to hate its own slop or we’re not as close to rsi auto-research as we though we were

ICLR rebuttal be like:

every advisor encouraging students on ICLR review day:

soon, new students will not know or remember how much the efficient googling and skimming of pdfs was a core skill in coursework and research.

our culture is being forgotten… 😢

Why you can’t do RSI as fast as you like - x.com/i/article/2098…

I wasn’t on the Paper and if I was, we would have been accepted WITHOUT NEEDING REVISIONS! Thank you for your attention to this matter.

i tell students good talks are like flying a plane where there’s
1) taking off - getting ppl curious/interested
2) flying - conveying the info
3) landing - giving ppl a memorable takeaway

many talks never take off. the audience just sit there, zone out and be bored. which is safe but ppl don’t feel like they got anything from it and forget you

even more dangerous is taking off, but not being able to fly and crashing (overpromising and then not delivering on what you promised and disappointing ppl). in which case ppl will actually be mad at you.

the best talks have to take off, fly, and land. which takes effort

Got a DM from a PR agency: congrats on the NeurIPS paper, would I like to stop by an AI company's booth and **post** about what they're doing?

Let me make sure I understand the deal. I spend my time at their booth, then use my name and my feed to market their product.

They pay for the booth, the swag, the sponsorship tier, and the agency that sent the message. The one part they don't pay for is the researcher whose post is supposed to make people care.

Said no, thanks.

Mechanistic interpretability was a huge missed opportunity. Many people wanted to join that effort, assuming it meant a mechanistic understanding of generalization. Instead, it was a bait and switch into an extremely niche, esoteric, and largely ill-conceived effort around understanding the roles of individual neurons. This has often been a challenge in alignment and interpretability: compelling initiatives that ought to be broadly defined, or at least mean what we would expect they mean, end up being restricted to a very narrow and esoteric set of topics. For example, if someone says they work on “interpretability” it does not mean what one would naively assume, e.g., interpretable methods, algorithms, representations, or model deconstruction, understanding the data, etc., but rather sparse autoencoders, influence functions, saliency maps. What a waste!

Dopamine does not spike highest for guaranteed rewards. It spikes highest at reward prediction error. If you want to have fun in life, you should stop avoiding risk and uncertainty.

Here’s 576GB worth of go f*ck yourself Sam Altman

“I’m not a betting man, but I have a prediction. It’s going to turn out that P=NP in some galactic sense, and all we’ll learn is that we spent fifty years answering the wrong question.”

The only thing necessary for the triumph of hype-farming is for the experts to say nothing~

Another banger from Ben Recht 🔥