Sugam05

Undergraduate | Eval Science for LLMs

GitHub: Vein05
cover art made by my dearest girlfriend

[Sugam05 @SugamPanthi]

New Paper!!
[1/4] Users paste text into a chat input, continue typing, and assume that the model can separate their afterthoughts from the given task ask. Models "absorb" casual afterthoughts as part of the instruction, and we instantiate the phenomena as absorption.

[Sugam05 @SugamPanthi]

@Ofirlin then, I'm once again sorry about the aggression from my side. I'll be more rational going forward, have a good day!

[Sugam05 @SugamPanthi]

why is it considered normal to upload peer reviewing materials to a 3rd party? YOU are no better than the ones you are trying to feel superior to!

[Ofir Lindenbaum @Ofirlin]

I checked the abstracts of 6 #ICLR submissions with @pangram. It flagged 5 as 100% AI-generated.

Some were also barely readable. I’m all for using AI to improve writing, but the result still needs careful editing.

Regardless of how accurate @pangram is, an abstract should clearly explain the problem, the contribution, and the findings. This is not the case for most abstracts I read this year @iclr_conf

[Sugam05 @SugamPanthi]

@Ofirlin My only problem with this is if you were the reviewer of those papers! I should have made that clear. If you're not the reviewer, I'm sorry for the aggression from my side!

Similarly,
https://t.co/BtRmFCrklb

[Sugam05 @SugamPanthi]

@CSProfKGD and then is goes

[Sugam05 @SugamPanthi]

@iarthsingh i highly belive most people still use gpt 3.5 or something, opus 5.5 would never make this mistake

[Sugam05 @SugamPanthi]

@CSProfKGD @ylecun is this part of the class videos you're talking about? looks great!

[Sugam05 @SugamPanthi]

@pyrodiscus i think it's mainly true for almost all niches, many don't even release code, or don't look into a particular attribute that might sink their method gain in the real world

[Sugam05 @SugamPanthi]

@ZAlvin39105 very cool work

[Sugam05 @SugamPanthi]

@CSProfKGD @phillip_isola thoughtful write up prof! such sad I'm not in computer vision :D

[Sugam05 @SugamPanthi]

I tried it, just like everything else, it serves no purpose.
The writing quality and the prose of an academic text is lost when using Emulate. As like many of it's predecessors, it tends to create unfaithful text compared to the original text provided. Good as a gimmic, but I'd rather good AI written document than a wrong "human" classified document.

[Agape Keleta @agapekeleta]

Introducing Emulate-1

A new undetectable AI model trained on how humans naturally write, removing AI slop and producing more authentic, human-sounding text.

I ran Emulate's outputs through Pangram, one of the stricter AI detectors. They came back as human nearly every time

[Sugam05 @SugamPanthi]

@CSProfKGD i think Claude has generous limits currently, but by the time you switch it's back to ChatGPT

[Sugam05 @SugamPanthi]

@mdredze everyone loves agents now!!

[Sugam05 @SugamPanthi]

@iarthsingh when you share :D I can get fairly good output with my finetuned olmo 32 b

[Sugam05 @SugamPanthi]

@CSProfKGD the hair is back!!! congrats!!

[Sugam05 @SugamPanthi]

@mar_kar_ i think many just want the early movers advantage on a new technology, and just generally don't care if what they produce is useful as a whole to the community or not

[Sugam05 @SugamPanthi]

@CSProfKGD please tell me i got phase 1, the portal is still 0 reviews submitted for me, hehe

[Sugam05 @SugamPanthi]

@CSProfKGD your hair got better the older you got it seems 🤣

goodluck1

[Sugam05 @SugamPanthi]

@iarthsingh hello arth, I look up to you very much! good luck!

[Sugam05 @SugamPanthi]

@Taodav yeah goota choose between being broke or super-duper-broke