Amine El Ouassouli

That it’s time-off o’clock.

X’s dominance ‘over’ as Bluesky becomes new hub for research Data indicates more scholars turning to alternative social media site to post about their work after Elon Musk’s Twitter takeover

'Bluesky has overtaken its flailing rival X in hosting posts related to new academic research, indicating the platform is fast becoming the go-to place for scholars to share their work.'

13 minutes of wisdom.

“No authorities in science”.

Amen to that.

ImageNet Moment for Reinforcement Learning? YouTube video by Machine Learning Street Talk

@jfoerst.bsky.social take on how the community sees the ARC Challenge and how we evaluate models and use benchmarks nowadays is 👌.

#more_science_less_hype (please).

PS: Amazing discussion and good brain food, as usual with MLST.

There is nothing truer than this true statement.

📍

bsky.app/profile/aelo...

I missed this one when it came out but I can tell that it is one of the most useful piece of research I’ve read in a while.

“GSM-Symbolic: Understanding the Limitations of Mathematical Reasoning in Large Language Models”

arxiv.org/html/2410.05...

Seeing “very successful” and “have a somewhat loose relationship with the truth” referring to the same people is what I can’t make sense of …

We really need better brain-power allocation. The current algorithm is kind of turning crazy.

The ToDo list: a revolution.

Est ce qu'on dirait que "la promotion à outrance des mathématiques est idéologique" ? ça laisse entendre que "l'IA" spécifiquement, à forte dose, est une sciences eugénisante par essence. On peut très bien en faire une interprétation/usage "progressiste" même si ce n'est pas dans l'air du temps.

À mon avis, ce type d’analyse n’aide absolument pas à se faire une opinion. Le « calcul » (parce que c’est ce que c’est au finale) n’est que le calcul. En faire quelque chose d’idéologique par essence est une sur-interprétation biaisée. L’ « IA » elle même ne porte rien du tout.

😭

That’s a very good one 👌🏽

Best: the most useful research you can do in the current context.
Worst: seems that it is not the main focus for now + maybe the pushbacks.

May the force be with you!

Is it an outlier, though?

(and one way of coping for me is to listen to MLST to hear more nuanced, or at sounder, views and opinions + reading)

100%

The more I read and listen to current debates in the field, the more I’m convinced that we have a model evaluation crisis.

I never understood people going to concerts to spend their time there attending through the tiny screens of their phones.

Basically, IMO, given that all assertions have different degrees of consensus in the population, accurate sequential token prediction may overlap or not with accurate “truth” in the “meaning” or conceptual realm.

The model may represent truthfully what’s in the dataset, even if it is untruthful. An analogy I often use: you don’t decide if evolution exists by popular vote. The vote tells what the population thinks. Research work even if it is coming from a single individual is more relevant in “truthfulness”

Is it just me or are we in an Eliza effect pandemic?

What is clear for me is that the current hype is not helping the calm development of these methods and collaboration with other fields.

PS: as someone mentioned, cross domain collaboration is key when it comes to ai research. It is hard, but it is key.

If it is not satisfactory at an epistemological level, it is not always clear at the moment and advances in the field will highlight that later. Is that non-integrity ? I would say no (maybe I’m mistaken).

Then, there is epistemology. What people call ai nowadays is inductive reasoning at a huge scale. It’s new, not mature (even for ai researchers), and, it’s WIP but, really promising. If people are using it, they are using approaches that are still being developed, thus, inherently experimental.

In my opinion, there are two layers in that question: an epistemological and a deantological one. If someone is using “ai” in a wrong way knowingly or for clearly bad reasons (e.g secure funding, for the hype), then yes we have an obvious integrity problem. That’s the deantological part.

Huh, what a year !

Happy new year, everyone ! May it be a better one than 2024 (it’s not that hard, though)

Take care of your loved ones.

PS: I was amazed by how many people use a p-value without fully understanding how to interpret it correctly. It doesn't make their work unacceptable, though. Everybody doesn't need to be a statistician.

Here is another one:

Do all neuroscientists understand entirely how an MRI works?

IMHO, it is more a matter of epistemologically sound interpretation—understanding what can be concluded and what can't from the results (and that is/should be the job of people making those AI systems/tools).