Raj Dabre
@prajdabre
Followers
18K
Following
10K
Media
1K
Statuses
10K
Senior Research Scientist - @google, Adjunct Faculty - @iitmadras, @iitbombay, Matchmaker - stealth Ex: @NICT_Publicity
Joined January 2012
Some disclaimers and life advice: 1. This is not an academic only account. I will post about various things like research, life advice, relationships and match making, jokes, memes, and the occasional rage bait with a disclaimer. 2. You will hear a lot of incessant babble on
6
1
74
So glad to see this getting launched, congratulations @MiteshKhapra and @ai4bharat team! 🎉 This arena fills an important gap and I encourage everyone to try it out. Glad @googlecloud and @GoogleDeepMind could provide initial support.
🚀 Announcing the Indic LLM-Arena 🇮🇳 At AI4Bharat (IIT Madras), our mission has always been clear - build open, inclusive, and world-class AI for Indian languages. To further this goal, today, we’re introducing the Indic LLM-Arena, a crowd-sourced, human-in-the-loop leaderboard
0
2
9
Why am I so obsessed with this photo? I've stared at it atleast 50 times over the past 24 hours.
14
0
59
Interesting paper. It basically shows that, in the middle layers, if you encourage MoE router behavior to route low-resource languages to experts which activate for English, the low-resource languages also benefit. A simple inference time intervention is enough to boost cross
3
1
29
This has been in the making for quite some time!
🚀 Announcing the Indic LLM-Arena 🇮🇳 At AI4Bharat (IIT Madras), our mission has always been clear - build open, inclusive, and world-class AI for Indian languages. To further this goal, today, we’re introducing the Indic LLM-Arena, a crowd-sourced, human-in-the-loop leaderboard
0
0
12
@DrAllyLouks The post in question:
0
0
2
My dear techie friends! Just an appeal. I am doing some really innovative AI viability survey research for @IndianSocietyAI, and we got some survey results, which triggered me to get more data to get enough validation. We crossed a threshold and are happy about that.
docs.google.com
Please note that this is an undergoing research effort by the Secretariat of the Indian Society of Artificial Intelligence and Law, in partnership with Indic Pacific Legal Research LLP, which is why...
1
4
6
it’s crazy how the same ideas seem to converge independently in the field at any point in time it’s like the ideas of a thousand brains being bottled up by a few select souls at the same time because it’s ripe
this weekend we are going to go back far far in the pre-historic times of ai (2019) to figure out if sequentiality and parallelism can be conjoined with the universal transformer.
8
9
130
How come none of y'all spotted Googe? I genuinely messed up. Updated now.
4
0
50
if you are an aspiring researcher here’s how you can embed yourself in the broader research community through jolly cooperation
@waefrebeorn @prajdabre here’s the recipe to find collaborators: - read research that interest you (multiple) - discuss with the researchers (usually first author) twitter/linkedin - post about your understanding of their research - try to push the research a bit further and post openly about it
3
3
37
A neat study on how to balance multilingal alignment with cultural retention.
Lots of work on cross-lingual alignment encourages multilingual LLMs to generalize knowledge across languages. But this push for uniformity creates a tension: what happens to knowledge that should remain local? We look into this trade-off of transfer and cultural erasure:🧵
1
0
17
🗺️ Are we making our #LLMs multilingual, or anglocentric? Much work brings languages closer to English, but that comes at the cost of crucial #cultural nuance. @h__j___han tackles this trade-off with surgical steering, adapting LLMs to cultural contexts at inference time.
Lots of work on cross-lingual alignment encourages multilingual LLMs to generalize knowledge across languages. But this push for uniformity creates a tension: what happens to knowledge that should remain local? We look into this trade-off of transfer and cultural erasure:🧵
0
10
50
Since most of you asked:
Ok since many of you asked here is some more info: 1. You can't become a tutor if you're a rando. They almost always accept IVY League students/graduates or the equivalent. Preference is for native English speaking people but flexible if you have top credentials. So unless you
1
7
177
Ok since many of you asked here is some more info: 1. You can't become a tutor if you're a rando. They almost always accept IVY League students/graduates or the equivalent. Preference is for native English speaking people but flexible if you have top credentials. So unless you
There's an obscure market involving tutoring rich kids from UAE and Saudi. You can work 10 hours a week and make as much money as an entry level SWE in FAANG.
17
53
1K
One of my favorite things about the knowledge distillation paper is a detail that most people probably miss out on, involving the scaling of the loss after KL divergence between the student and the teacher model. This one scaling factor is needed to ensure that your distillation
7
7
126