←
rohan4,738 posts
rohan avatar

rohan

@lets_dig_deeper

founder & ceo @rumik_ai,building @iradotso | my balls have led me to places i won't go with a gun

◷ Joined November 2018▧ 683 Photos and videos
rohan avatar
rohan@lets_dig_deeper·Sep 8

introducing rumik oss 1 india's first state of the art open source, fully pre-trained voice model capable of speaking 20+ languages trained on only 66k hours of data, it beats even the top closed models on multiple benchmarks including emotion realism weights are now live 👇🏻

2:05
rohan avatar
rohan@lets_dig_deeper·8h

❤️

ok been using ira for a few weeks and it's lowkey better than most of the ai tools i've tried the proactivity is what gets me. it just pings me about stuff before i even think of it @rumik_ai @lets_dig_deeper y'all are cooking!!

andy avatar
⟳ rohan retweeted
andy@prompt_Tunes·Nov 17

you can get (somewhat) rid of the causal dependency between the codebook estimation for the nth codebook on the prev 1,2 ... n-1 codebooks in your standard RVQ setting by having a sort of global-local-global design while doing feature extraction from the incoming signal (1/n)

Post attachment

(2/n) the key difference is that it is not an autoencoder. most codecs, Encodec, DAC, Mimi minimize the reconstruction loss for |audio - reconstructed_audio|. here the entire learning happens in latent space (much like JEPA). the encoder learns by predicting masked audio in latent space. then a separate gan handles reconstruction. the architecture in one picture: audio → jepa style encoder → q2d2 geometry aware quantizer → hifi-gan decoder → audio 12.5 Hz frame rate (80 ms/frame) 128-dim latent space

Post attachment
ira avatar
⟳ rohan retweeted
ira@iradotso·Nov 17

We are resuming with few invites, dm for access

ira avatar
⟳ rohan retweeted
ira@iradotso·Nov 16

You can comments for early access for @iradotso, we have resumed invites.

rohan avatar
rohan@lets_dig_deeper·Nov 16

1. yes 2. yes 3. yes

3 key questions for every consumer AI product: 1. Does your product give people new superpowers? 2. Does your product get better as more people use it? 3. Does your product get better when models get better?

andy avatar
⟳ rohan retweeted
andy@prompt_Tunes·Nov 16

i had to make some stability fixes with some grad clipping for batch A's gradient, but this seems to work.

Post attachment

Compute grows fast but pretraining corpora don’t, so data-constrained optimization matters! We propose Training-Time Testing (T³): each step takes a trial step on half the batch, tests it on the other half, and updates with the meta-gradient. It achieves SOTA on SlowRun!

0:30
rohan avatar
rohan@lets_dig_deeper·Nov 15

indeed, exactly what i'm seeing upclose

With agentic commerce, the biggest loser is obviously Google Product discovery will all move to agents. The traditional 10 blue links are dead. What an amazing run. I had no idea how it would be defeated, but agents are just a 10x better search experience. Google is presenting no viable consumer agents. Maybe their best bet is to acquire Instinct but that’ll cost like $30b now.

ira avatar
⟳ rohan retweeted
ira@iradotso·Nov 15

We are resuming with few invites!!

ira avatar
⟳ rohan retweeted
ira@iradotso·Nov 15

Dropped some ios invites, check your dms.