Stephane Deny
@StphTphsn1
Followers
4K
Following
9K
Statuses
4K
Neuroscience, ML, also other things.
Joined April 2013
@andrewgwils One of the main principles of Wikipedia is to provide sources of information. On the other hand, LLMs carefully conceal their (copyrighted) sources through a thin layer of “intelligence” / rewording. One would replace the other at a great loss for the user 🤷♂️
0
0
3
RT @MLStreetTalk: Professor @randall_balestr discussing some exciting research he has been working on recently, in particular around spline…
0
31
0
RT @JeanRemiKing: Two new studies from our team we're particularly happy about Study 1: Brain-to-Text Decoding:
0
67
0
@giffmana @roydanroy (1) the public largely funds academic science, so they do deserve the utmost respect from us (2) the public not trusting the way science is currently done does not mean they don’t believe in the importance of science or facts, these two things often conflated
1
0
3
RT @StphTphsn1: "A comparison between humans and AI at recognizing objects in unusual poses" now published at TMLR! 🔗
0
7
0
"A comparison between humans and AI at recognizing objects in unusual poses" now published at TMLR! 🔗 We thank the editors and reviewers of TMLR for a smooth and fair reviewing process!
What do you see in this image? Deep learning has closed the gap with human vision on many benchmarks. But does it mean that deep nets are *as robust* as humans in rare scenarii, such as when objects are shown in unusual poses? We explore this question: 🧵
0
7
11
@banburismus_ @leedsharkey Thanks for the great forensic work. Any take about the 3rd offender ? (twitter has often be full of loud and inaccurate takes and profound 3likes takes i don’t think it changed that much)
0
0
1
0
0
0
@3_deame @numerounochef @ChombaBupe If we cannot agree on a statement being relatively correct compared to alternatives, then I don't think we can say much, and this debate is void as well. But, yeah, as a scientist, I strive for correctness in my reasoning and conclusions, although sure it is not fashionable
1
0
1
@3_deame @numerounochef @ChombaBupe Got it but for me these are for me two separate questions (free will and reasoning). Usually reasoning is constrained by a logic, assumptions and observations and there is a single correct answer (correct reasoning), it's not really the best example of an exercise in free will.
1
0
3
Currently trying to replicate (or fail to replicate) the "SAEs can interpret randomly initialized transformers" result on SmolLM2 135M, which was trained on 2T high quality tokens. Their paper used Pythia Fraction of variance unexplained is much higher for random than trained
0
0
1
1
0
0
@3_deame @numerounochef @ChombaBupe For every flaw pointed at regarding LLMs, there is always this reply that humans can do the same mistakes. Usually, that's not true of *attentive* humans. Attentive humans do not do the "same mistakes".
2
0
2