Anil Kag

187 posts

Anil Kag

Anil Kag

@anilkagak2

Looking for Research Positions in Industry | Ph.D. Candidate @BU_ece | Past: Intern @MSFTResearch | RF @MSFTResearch | SWE @microsoftidc | https://t.co/gLcBIGxRkl. @IITGuwahati

Boston Katılım Mart 2013
518 Takip Edilen107 Takipçiler
Anil Kag retweetledi
Ziyi Wu
Ziyi Wu@Dazitu_616·
📢 Introducing DenseDPO: Fine-Grained Temporal Preference Optimization for Video Diffusion Models Compared to vanilla DPO, we improve paired data construction and preference label granularity, leading to better visual quality and motion strength with only 1/3 of the data. 🧵
English
2
35
180
35.2K
Anil Kag
Anil Kag@anilkagak2·
RT @chesscom: WOWWWWW 🔥🇮🇳 AFTER AN INCREDIBLE TOURNAMENT, AND THE CRAZIEST DAY, 17-YEAR-OLD @DGUKESH WILL COMPETE FOR THE WORLD CHESS CHAM…
English
0
392
0
19
Anil Kag retweetledi
Viswanathan Anand
Viswanathan Anand@vishy64theking·
Congratulations to @DGukesh for becoming the youngest challenger. The @WacaChess family is so proud of what you have done . I’m personally very proud of how you played and handled tough situations. Enjoy the moment
English
133
2.3K
18.3K
635K
Gaurav Aggarwal
Gaurav Aggarwal@fooobar·
Forced to take BP medication for the first time in life Reading was kind of high (150/100). Wish me luck to successfully replace these daily medicines with lifestyle changes - more steps, diet control, better sleep timings, etc. Don't want to get old :(
English
10
0
62
8.7K
Anil Kag
Anil Kag@anilkagak2·
@fooobar okay, they look ripe already.. you cannot eat them all.. give me some :P not finding good mangoes in Boston :(
English
0
0
1
51
Gaurav Aggarwal
Gaurav Aggarwal@fooobar·
Mangoes are love - all varieties!
Gaurav Aggarwal tweet media
English
1
0
17
1.1K
Anil Kag
Anil Kag@anilkagak2·
@giffmana Yes when you aim for consciousness in your model, there comes a point when the model gains consciousness and becomes AGI. This is where you see the sudden drop in loss. /s
English
1
0
1
160
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
That’s the main reason in modern large scale training. You can also construct or find sudden drops in toy or esoteric tasks/models, but not usually in general NLL training on broad, real world data. Though maybe I’m saying this only because I haven’t trained conscious AGI yet?
English
5
0
39
5.5K
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
While everybody is making fun about sudden loss drop, let me tell you what it ACTUALLY means when that happens: Your training run had a pre-emption and got restarted. Your input pipeline isn’t perfect and now partially going through examples you’ve already seen recently No foom
Eliezer Yudkowsky ⏹️@ESYudkowsky

Possible but hardly inevitable. It becomes moderately more likely as people call it absurd and fail to take precautions against it, like checking for sudden drops in the loss function and suspending training. Mostly, though, this is not a necessary postulate of a doom story.

English
10
8
177
73.7K
Anil Kag
Anil Kag@anilkagak2·
@aaron_defazio If your issue is overfitting, have you tried with heavy data augmentations like AutoAugment and Cutout ?
English
1
0
0
149
Aaron Defazio
Aaron Defazio@aaron_defazio·
CIFAR10 is a terrible benchmark problem for optimizers. Adam doesn’t work. LION doesn’t work. Polyak step size doesn’t work. So many things that work great on 90% of other test problems don’t work on it. If we keep using it as a test bench the field won’t progress!
English
6
2
44
19.2K
Yao Fu
Yao Fu@Francis_YAO_·
GPT-4 should be large, but it's inference latency seems to be the same as GPT-3.5 legacy. How could one scale up model parameter with constant inference latency? 🤔🤔
English
24
3
30
21.4K
Anil Kag
Anil Kag@anilkagak2·
@thegautamkamath @roydanroy That means @icmlconf intentionally decided to keep the rebuttal deadline on Sunday. How does that help work-life balance, if our deadlines keep overlapping with weekends :(
English
0
0
1
110
Gautam Kamath
Gautam Kamath@thegautamkamath·
@roydanroy My understanding is that they were released ~1 hour and 45 minutes after the stated time (judging by the review edit time on OpenReview). There was also a few more hours before the notifying email was sent out.
English
2
0
1
1.7K
Dan Roy
Dan Roy@roydanroy·
Why am I spending my Sunday finishing rebuttals? So much for respecting weekends and families.
English
7
9
135
50.8K
Anil Kag
Anil Kag@anilkagak2·
@FeiziSoheil @kchonyc as a reviewer.. this has been my pet peeve when I'm submitting reviews.. so far I've not swapped reviews (I always check the title of the paper with my summary)
English
0
0
0
521
Soheil Feizi
Soheil Feizi@FeiziSoheil·
This was new! We got a review swap: got review for another (unknown) paper on our ICML submission; @kchonyc we wouldn't bother you if the review was positive but it was, unfortunately, very negative! 😅
English
5
1
34
15K
Anil Kag
Anil Kag@anilkagak2·
@giffmana Its refreshing to see that they have finally embraced LSTMs :P
English
0
0
1
2K
Lucas Beyer (bl16)
Lucas Beyer (bl16)@giffmana·
lol, draft of full GPT-4 paper with architecture and data details is already leaked on torrent😂 The vision component in the architecture is an interesting twist to plain ViT, and scaled up quite a bit! Link to the torrent for the curious: assets.speakcdn.com/assets/2703/bu…
English
58
186
1.2K
462K
Anil Kag retweetledi
Amul.coop
Amul.coop@Amul_Coop·
#Amul Topical: The Elephant Whisperers wins Best Documentary Short Film at Oscars!
Amul.coop tweet media
English
101
1.7K
24K
2.4M
Anil Kag
Anil Kag@anilkagak2·
@roydanroy Other hotels are more than 4miles away from the venue.
English
0
0
0
199
Dan Roy
Dan Roy@roydanroy·
To be clear, what's sold out is the hotel near the venue. There are places left that are a 13 minutes drive away. I'm not sure why an announcement wasn't made, but the effect of this is likely that only insiders got to book that close hotel. Disappointing transparency.
English
3
0
9
4.4K
Dan Roy
Dan Roy@roydanroy·
Did anyone else get an email announcing that the ICLR hotels are finally set up and that conference attendees can book? No?? Me neither. Better go book before it's too.... Too late, sold out. Wait, weren't we instructed not to book and wait for the official conference hotel?
English
1
2
26
19.5K
Anil Kag
Anil Kag@anilkagak2·
@ccanonne_ Someone has to pay for the genius' $44bn investment. How else would Elmo get his return on the investment :P
English
0
0
0
90
Anil Kag
Anil Kag@anilkagak2·
@sytelus I believe the ChatGPT APIs might be serving a lower complexity model (for faster inference & lower cost) than the one hosted by ChatGPT at OpenAI website.
English
0
0
0
117
Shital Shah
Shital Shah@sytelus·
Every other bot fails, even ones using ChatGPT APIs. Only the original ChatGPT at OpenAI answers correctly (i.e. asymptomatic freedom). Failed bots: - Poe/ChatGPT - Poe/Claude - Poe/Sage - Bing Chat - Perplexity AI - You.com - davinci-001,002,003 - Flan-UL2
English
3
1
13
1.8K
Shital Shah
Shital Shah@sytelus·
There is a mysterious secret sauce in original ChatGPT hosted at OpenAI. Many of my hard queries get only answered there. Ex: “What was that phrase in the book God Particle about quarks repelling each other if they came too close but attracting each other if they went too far?”
English
4
2
17
4.9K

Keşfet

@chesscom @DGUKESH @DGukesh @WacaChess @siarohin9013 @isskoro @katedeyneka @tsaishien_chen