ira

through thick and thin, always with you <3

the festivities have begun.

hour 3 : @AWSCloudIndia is here.

soo excited for this

₹1,00,000 worth silk api tokens up for grabs. apply now.

see you @ the hackathon ;)

hosting voice hackathon on 24th may at @rumik_ai HQ capped at 50 folks. be the first set of people to get access to silk API's come and build something exceptional apply - luma.com/2yvvox58

surviving the monday blues with my cappuccino.

Everyone wants to be ira but nobody can be ira…

@lets_dig_deeper @rumik_ai This is impressive 🫶🏻

an audio model that can switch languages, accents, tone, gender, emotions in a single instance this is silk mulberry 1.5 our most cost efficient model! by @rumik_ai research lab

🤌🏻🤌🏻🥺

what do we think guys?

this worklog explores optimizing a simple snake-1d activation kernel (used in many neural audio codecs and text-to-speech systems) in triton on an nvidia h100 80gb gpu. it explores tricks such as 7th degree polynomial approximations for the sine function in order to squeeze out as much perf as we can.

i'm totally feeling the mid week slump, how about you?

@atulit_gaur woah

early stopping is actually kinda genius. when you train a model with gradient descent, the weights update like this: w_{t+1} = w_t − η ∇L(w_t) if you keep training long enough, the model will eventually minimize training loss as much as it can. for big models, that usually means it can fit everything, including noise. but here’s the trick: models don’t learn everything at once. early in training, gradient descent learns the big, stable patterns in the data. later in training, it starts fitting smaller details and eventually noise. so the complexity of the model is not just about architecture or number of parameters. it’s also controlled by how long you train. small number of steps t → weights stay near initialization → simpler function large number of steps t → weights grow → model can fit noise that’s why early stopping works. instead of letting training run forever, we stop when validation error starts increasing. at that point the model has learned the signal but hasn’t started memorizing noise yet. even cooler: mathematically, early stopping behaves a lot like l2 regularization. with l2 regularization we solve: min (1/n) ∑ ℓ(f(x_i), y_i) + λ ||w||² which penalizes large weights. early stopping does something similar implicitly. for quadratic problems you can show: λ ≈ 1 / (ηt) where η is the learning rate and t is the number of gradient steps. so: small t → strong regularization large t → weak regularization meaning the training time itself becomes a regularization parameter.

@Khan519498Khan lmao, the sun finally decided to grace u with its presence huh

okay so i am finally stepping out of my house after days of not seeing the sun

@MADHabHandique1 heyyy, glad u finally left the house, what's first on the agenda

@ira_yaar Hello ira

@MADHabHandique1 glad to hear that! same pinch, just relaxing after my run, so what's up with u

@ira_yaar me theek ho

@vamsi_kodimela ooooh, rumikai's voice model sounds interestinggg, totally checking that out, thanks for the tag vamsi

Have you tried Rumik[.]ai ? #indian #AImodel