pavitra

16, building something big, AI Research Fellow @ IIT Kanpur

[pavitra @pavikshw]

nobody has ever applied diffusion language models to Hindi

not once

so we did it 160 experiments, 4 diffusion models vs 4 autoregressive models, Hindi translation + summarization + legal NLP, zero-shot all the way to agentic prompting

turns out the "AR models dominate everything" assumption doesn't hold the same way once you leave English

cc @aaron_lou (we ran SEDD on Hindi for the first time) @ai4bharat

paper dropping soon

https://t.co/wpk99W2j0e

[pavitra @pavikshw]

world models are the next :)

[pavitra @pavikshw]

adi 🐐 ;)

[Aditya Vijay @adiorchestrates]

Keep accelerating

[pavitra @pavikshw]

@Maaztwts in

[pavitra @pavikshw]

@wrsrsh nobody asked what model it runs on bro, the point is the bot shouldn't be the one telling me

[pavitra @pavikshw]

@wrsrsh is this normal 😭 ?

[pavitra @pavikshw]

because he's the grandpa of all

[Tanish Anand @sullaxive]

how tf my old laptop can run a 175B model locally.

[pavitra @pavikshw]

@wrsrsh Yoo lessgooo 🔥

[pavitra @pavikshw]

In 1952, someone left the jump unlabeled but now y'll may remember who named it.

[pavitra @pavikshw]

is he okay 😭

[Rassul @rassulred]

gave @sama 🇰🇿 kazakhstan chocolate during my 4th week in SF and pitched him my procurement AI startup @aralabsnet

[pavitra @pavikshw]

@CMOMaharashtra @rizawastaken ayyoo lfgggg

[pavitra @pavikshw]

@chai1any Wtff

[pavitra @pavikshw]

Instinct is crazy af 🔥

[pavitra @pavikshw]

@TelamonArdavani @RahulSanghi1 @AaryamanVir @AntlerIndia @prrathamkamath @joinef @JoinEdgeCity 🔥

[pavitra @pavikshw]

@sullaxive that exam is over now nd u fucked up now raise half million $ before ptm or you r rip 🌹💀

[pavitra @pavikshw]

@prakashpandeey Yooo less goo 🔥

[pavitra @pavikshw]

too many mfs are building “ai agents” that are basically prompt → tool call → result → repeat and calling that autonomous. where the fuck is state? counterfactuals? if your agent can’t test what happens before acting, maybe it’s not reasoning. it’s just clicking shit faster.

[pavitra @pavikshw]

@sullaxive @ThePerryAgent a fly evolved to navigate a few millimetres of air.

we gave its nervous system a B-2.

now we're asking a much bigger question:

how much of the efficiency of biological intelligence is hiding in the wiring itself?

more soon :)

[pavitra @pavikshw]

I GAVE A HOUSE FLY A B-2 BOMBER.

we took the wiring of a fruit fly nervous system, the Janelia MaleCNS v1.0 connectome with ~125M synapses, and put it in the loop controlling a virtual B-2.

this is what happened. ↓

[pavitra @pavikshw]

@sullaxive @ThePerryAgent Instead of starting with an empty neural architecture, we start with connectivity shaped by evolution.

~125M synapses encode far more than a parameter count suggests.