Gal Katzh retweeted
Iโm joining the Weizmann Institute as an Assistant Professor
(known[?] as โSenior Scientistโ ๐งโ๐ฌ) and starting CoLab๐ถ!
We do pretraining, agents and evaluation the way science should be done:
in a community, together.
Come research with us ๐
discord.gg/gVeGEjWTtv
Gal Katzh retweeted
ื ืชื ืืื ืืขืช ืืงืืคืืื "ืืขืจืืื ืื ืื ืืช ืืืืืจืืช".
ืืืืจื ืืืืฆืจ ืื ืืืืืืืฆืื ืืืืืจืืช, ืขืืจืื ืืขื ืืื ืจืืฉ ืฉื"ื ืืื ื, ืืคื"ื ืืืฉืืจื ืื ื ืืื ืืื"ืจ ืืขืืช ืืืืืจืืช ืกืืืืจื.
ืืช ืืจืืืืช ืืืืืฃ ืืืืืื ืฉื ืืืืืจืืช, ืืกืจ ืฉืืื ืงืืืืืจ โ "ืื ืื ืืืืจ ืืืืืจืืช ืืืืืจ ืืขืจืื ืืืขื ืืืืืื".
ืื ืืืืฉ ืขื ืืชืืืจ ืืืฉืื?
ืงืืืืืจ ืืื ืื ืืืฉ ืืืง ืื ืฉืืคื ืืืืืืก.
ืืืืืจ ืืงืืคืืื ืจ ืฉืืชืืื ืืคืขืืืืช ืืืืืื ืืื ืืืืืืืฆืื ืืืืืจืืช ืืืืจื ืืขืจืืืช.
ืืฆืืืื ืืืกื: ืืฉืจื ืืงืืคืืื ืื "ืงืืืืืจ ืขื ืืจ" ืืชืืื ืืืฆืืืชื ืืืืขืืจ ืืช ืืฉืื ืืงืืคืืื ืืืืืจืืช ื-2019
Gal Katzh retweeted
ืืชื ืคืืืช ืฉืืืืื ืืืจืฅ ืืฉืจืื ื ืืืจื ืืืื. ืืจืืืช ืืกืืคืืจืื ืขื ืื ืขื ืืื ืชืืืจืื-ื ืืกืขืื ืืื ื ืืืืกืกืื ืืื ืขื ืืืืช ืืจืืื ืืืงื ืืฉืื, ืืชืืฆืืชื ืืืืืื ืืฆืืืฆืืช ืืืงืฉืืฉ ื ืืื ืืืฉืืืฉื ืืื ืฉื ืืืฉืืจ [ืืืืจ ืืืื ืฉืื ืืื ืขืืืจ ืืชืืืจืื].
ืกืืฃ ืื ืกืืฃ
Today @GoogleMaps is getting its biggest upgrade in over a decade. By combining our Gemini models with a deep understanding of the world, Maps now unlocks entirely new possibilities for how you navigate and explore. Hereโs what you need to know ๐งต
98
awesome history-of-math-words:
mathshistory.st-andrews.ac.uโฆ
h/t @MoritzSchauer
2
429
Inspiring recap for the LLM jailbreaking competition (winning defense team!)
My team with @Yuvlem won๐ฅfirst place in the defense track of the Capture The Flag-LLM Challenge at @satml_conf. In the competition, each team hides a secret in an LLM context, and other teams try to discover it. Here's how we used ideas from Judo and Cooking to do it๐
๐งต1/11
1
219
Fascinating work by the brilliant @eladgranot and my PI, illuminating the importance of element-wise activation functions, using a powerful, little known technique for yielding the upper bound
Thrilled to share our paper with Amit Daniely at the ALT 2024 conference! We prove a tight generalization bound for two-layer networks, shedding light on the role of the element-wise property of activation functions in achieving these bounds. โจ #ALT2024
arxiv.org/abs/2211.09634
3
204
Gal Katzh retweeted
Can we design good transaction fee mechanisms (TFMs)?
If so, what are the trade-offs?
Together with @yaish_aviv, we show an impossibility for deterministic TFMs, and limitations for randomized TFMs.
Paper: doi.org/10.48550/arXiv.2402.โฆ
Read on for more details (1/10)
Gal Katzh retweeted
Wow!
This might be the most interesting detail about Palestinian conflict deaths: they're basically just male.
Israel must not be indiscriminate.
Replying to @cremieuxrecueil
There's also data on the Israeli age distribution (they don't specify whether or not victims took part in hostilities), so I split it up by gender. I'll include the Palestinian one split by Gender as well for comparison.
Gal Katzh retweeted
Did you think your model is protected if you publish it as a black box? Think twice.
Happy to share our #ICLR2023 paper:
An Exact Poly-Time Membership-Queries Algorithm for Extraction a three-Layer ReLU Network
arxiv.org/abs/2105.09673
w. Amit Daniely
Gal Katzh retweeted
We want to pretrain๐ค
Instead we finetune๐ฎ๐
Could we collaborate?๐ค
ColD Fusion:
๐Recycle finetuning to multitask
โก๏ธevolve pretrained models forever
On 35 datasets
+2% improvement over RoBERTa
+7% in few shot settings
๐งต
#NLProc #MachinLearning #NLP #ML #modelRecyclying
Gal Katzh retweeted
This meme implies weโre living in uniquely stupid times, this 1994 letter to an editor begs to differ:
Proud to be a data point in this cool research
Analogies are everywhere!
We Gotta Catch 'Em All! โ๐
๐ฃ #emnlp2022 (main conference) long paper ๐:
๐ช Life is a Circus and We are the Clowns: Automatically Finding Analogies between Situations and Processes ๐คก
w/ @HyadataLab
paper: arxiv.org/abs/2210.12197
๐งต 1/
Gal Katzh retweeted
Transformers are not Seq2Seq
Given a context
They predict a single token
We use this to update representations between predicted tokens
& feed a changing graph relevant to the current token
arxiv.org/abs/2101.12640
me @oabend
#conll2020 #NLProc
A few hours ago I had a discussion with friends about the state of ethics in CS academy. I was on the defending side. Seeing a leading prof dismissing such a burning ethical issue (which is a gray area imho), I realize I might have been completely wrong.
Replying to @Grady_Booch
"I don't want my Image to be indexed."
Then.
Don't.
Post.
Your.
Image.
(Or block the bot.)





























