Prashant Shishodia
@pshishodiaa
Followers
312
Following
1K
Media
52
Statuses
275
building frontier of speech models @KalpaLabsAI (yc f25) ex-Action Models @GoogleAssistant
sf
Joined June 2023
sf food so bland, i sometimes open zomato just to feel something
3
0
9
Tensr (@TensrLabs) is building fully autonomous robotic factories so hardware teams can spin up production the way developers spin up cloud compute.
16
16
100
found a new emergent capability in our models today that's so insane my jaw is on the floor. I don't think anyone can predict this. bowing down to the optimization gods.
2
0
8
when we started scaling to larger clusters everything started breaking unexpectedly (it deserves it's own separate blog post). It wouldn't have been possible without @SfTensor's incredible support even as late as 3 am at night. @ycombinator is truly a magical place.
Really excited to see this one launch in part because its a great story of two YC teams working together. Limited in GPU capacity and rushing to get their model trained, this team worked with another YC company in the batch working on GPU software and infrastructure until 3am for
1
0
8
Excited to have powered the training infrastructure for @KalpaLabsAI's latest model. Congrats to @JhaGauti and @pshishodiaa on the launch and excited to see where you take it!
Kalpa Labs (@KalpaLabsAI) is training one generalist model for all speech tasks. Steer it like an LLM, give long system prompts & make use of native contextual awareness. Congrats on the launch, @pshishodiaa and @jhagauti! https://t.co/V0AbssT7CN
1
3
9
this is just the beginning. The lightning strike of bitter lesson for speech models is not very far.
Kalpa Labs (@KalpaLabsAI) is training one generalist model for all speech tasks. Steer it like an LLM, give long system prompts & make use of native contextual awareness. Congrats on the launch, @pshishodiaa and @jhagauti! https://t.co/V0AbssT7CN
4
0
10
scaling bet paid off. but the ladder is too long, and we've a lot to climb.
0
0
0
messaged a billionaire & he replied back in 6 mins????????????
2
0
7
git, uv, ncdu, aria2c - the four horsemen of goated software
1
0
1
tinker's hparam estimate: claims < 0.5% regret compared to full hparam sweep.
0
0
1
minor detail that i only realized today: 1 / sqrt(fan_in) not only ensures std(output_activations) = std(input_activations) but for embedding, initializing with 1/root(codebook_dim) also ensures norm(codebook[i]) = 1 🤯
0
1
2
TIL llama3 had inbuilt support for speech, but they never released it.
0
0
1
one of the most unexpected surprises of moving to sf from blr was the realisation that blr undeniably had significantly higher talent density than sf. sf still wins because of grit, ambition & culture.
0
0
2
Meesho's cloud bill is $70M / year ????????
Amazon vs Meesho 🥊 Amazon's cloud unit, AWS, states that Meesho has not paid bills worth Rs 127 crore. Meesho has said that Amazon’s services were deficient and made a counter-claim of Rs 86 cr, due to loss of business Since then, Meesho has moved from AWS to Google Cloud 👋
0
0
1
fine-tuning large models doesn't really makes so much sense because ICL and few shots are already good enough & it's not even cost / latency effective for an enterprise to use it across 100 different teams / features which is really what's in the box for democratisation of
0
0
1