Yassine Taoudi-Benchekroun

I don’t see LeCun claiming he invented dropout

Funny to see how quickly everyone is, without any remorse, switching to DeepSeek R1 over OpenAI o1. In this world of hyper-quick product shipping, there is no such thing as customer loyalty. OpenAI's sketchy reputation surely doesn't help.

This is from Richard Hamming’s « The Art of Doing Science and Engineering », written in 1997.

That’s about 20 years before Richard Sutton’s « The Bitter Lesson ».

I’m sure Schmidhuber might have something to say about this lol.

Today I heard a friend saying that he wanted to start writing, but is waiting to find something interesting to write about.

This is the same as saying:

"I will start playing piano when I stumble upon a good idea for a tune".

Start writing today.

Friends, what are your best articles/resources on how to write and story-tell effectively?

Note: this is for technical/scientific writing, not prose.

Today I found out that a friend of mine got a huge $$$ contract out of a conversation he had at my birthday party.

Parties are useful. Party more, and organize more parties.

The papers I admire most introduce such breakthrough ideas that they barely need to justify their relevance. They’re almost always very short.

The results speak for themselves. Even more so, the confidence of the authors, who don’t bother over-justifying their ideas, speaks even louder.

It's scary how increasingly difficult it's become for academia to produce useful and impactful findings in terms of frontier AI research.

f***-me compute, f***-me data, fast and high-quality *engineering* are what's pushing the limits.

"It's engineering then, not science!".

Is it really?

"This is not merely incremental improvement, but a genuine breakthrough".

System 1 is not the bottleneck anymore.

Now we ought to work on getting 'test-time-adaptation' and 'search' done, with *resource constraints*.

Every schmidhuber paper is a piece of art. I also am annoyed at all the non-scientific and egocentric nonsense, but I gotta give this to him. He writes prophetic papers.

Hi NeurIPS!

Explore ~4,500 NeurIPS papers in this interactive visualization:

jalammar.github.io/assets/neuri...
(Click on a point to see the paper on the website)

Uses @cohere.com models and @lelandmcinnes.bsky.social's datamapplot/umap to help make sense of the overwhelming scale of NeurIPS.

I understood #BlueSky was actually great when I opened it this morning and did not see a single tweet about #LuigiMangione.

Not saying that this is not important (it is), but I understood this app really tries to give you the feed you want.

Reading #MoneyStuff by @matt-levine.bsky.social is such a strange blend of “why is he explaining this like I have 2.5 brain cells” and “fuck I just understood a crazy subtle intricacy in the market that is so niche and juicy”. I love it.

Test-time-training will lead to major rethinking of how we do AI.

Soon enough, we will start seeing anti-AI political movements.

Yassine Taoudi-Benchekroun's review of A Little Life 5/5: After several days of processing this heart-wrenching book, I found myself needing to make a choice about what to take away from it. While it might be most obvious to obsess over the unspeakable ...

My Review of "A Little Life"

www.goodreads.com/review/show/...

Am I the only one very puzzled by this claim? Associating hallucination to generalization in #LLM is just bs and I don't get how this went through #Nature peer review.

From "LLMs surpass human experts in predicting neuroscience results " in Nature Human Behaviour www.nature.com/articles/s41...

Attended a talk from a Science editor today:

"I won't tell you that you should spend 10k to hire third parties to help with you with your figures... But it helps".

Yeah, that's where we've gotten to.