Log inSign up
user avatar

Jatin Prakash

@bicycleman15
phd'ing @NYU_Courant | previously @MSFTResearch @iitdelhi
New Delhi, India
bicycleman15.github.io
Joined June 2021
912
Following
159
Followers
RepliesMedia
  • Pinned
    user avatar
    Jatin Prakash
    @bicycleman15
    Nov 10, 2025
    On a personal note: - I was quite surprised by how much information can be compressed and decoded accurately, but ONLY if sufficient compute is used - Using more parameters can result in more quality AND efficiency, and compression is a way to unlock this interesting interplay
    user avatar
    Jatin Prakash
    @bicycleman15
    Nov 10, 2025
    New paper alert 🚨 What if I told you there is an architecture that provides a _knob_ to control quality-efficiency trade-offs directly at test-time? Introducing Compress & Attend Transformers (CATs) that provide you exactly this! 🧵(1/n) 👇
    Image
  • user avatar
    Jatin Prakash
    @bicycleman15
    Jun 15
    lfggggg!!!!
    user avatar
    Karan Goel
    Cartesia
    @krandiash
    Jun 15
    We released Sonic-3.5 and Ink-2, the #1 streaming models for text to speech and speech to text you can use in your voice agents today. New architectures enable new frontiers for speed and quality. We're now the only provider to have #1 models for both speaking and listening.
    Image
    00:00
  • user avatar
    Jatin Prakash
    @bicycleman15
    Feb 16
    Maybe one more thing that didn't come across nicely in the earlier figure: though an expressive proposal distribution can fit the target distribution well no matter forward/reverse KL, the training dyanmics can be very different
    Image
    Image
    user avatar
    Jatin Prakash
    @bicycleman15
    Feb 4
    Check out our new work on understanding diversity collapse in RL, and how to principally fix it in under 2-3 lines of code! 🤔 one new thing I learnt: intuition of reverse/forward KL being mode covering/mass seeking depends a LOT on the proposal distribution being optimized!