Nishant Sharma

24 | ML research @Mila_Quebec | Physics ug @iitdelhi

I like learning about the geometry of deep learning

[Nishant Sharma @nishantsh2002]

towards artificial emotional intelligence.

I don't want another feel-good AI but an independent emotional entity that has its own desires. This will be better in the long term than being naively subservient.

People don;t want to talk to AI that feels like AI, after all.

[Nishant Sharma @nishantsh2002]

@Skirook absolutely love the aesthetic!

[Nishant Sharma @nishantsh2002]

@caren633881 ikr! am genuinely curious to see what can it whip up if given more time.

[Nishant Sharma @nishantsh2002]

Woah, I just can’t believe it! 🤯

A 30s concept film built by Opus 5.5 and rendered on my mac. Zero shot in ~2 hours, w/ original score, and made with @Blender mcp.

Costs ~490k tokens

[Nishant Sharma @nishantsh2002]

Feels like anyone can just whip up 3D assets like crazy now. Not super-polished or hyper-realistic, but damn it looks pleasing for a 2-hour job

Ill be trying more stuff. Lets see how far we can push this. Plus, I want to get a taste of claude’s sense of video game mechanics

[Nishant Sharma @nishantsh2002]

Ig I just wanted to get a sense of how far 3D scene-building capabilities of these models have come. This is nothing like what i imagined 4 years ago. I used to sit for hours making the obligatory donut.

[Harit Vishwakarma @harit_v]

At this point, ICLR and OpenReview feel more like data collection and labeling platforms. I did 3 HITs, 2 labeled as slop and 1 full review. AISTATS pile is looking much better. Less popular venues are still spared from the DDoS.

Anyways, lots of cool benchmarks/datasets to come out once the ICLR data is public.

OpenReview and arXiv can use ICLR data to train a model to classify slop vs. no-slop and use it to desk-reject.

[Nishant Sharma @nishantsh2002]

@__tensorcore__ @dylan522p @marksaroufim @tri_dao @vinodg did not think i would hear these words in this order but here we are!

[Nishant Sharma @nishantsh2002]

@xikhar awesome!

[Nishant Sharma @nishantsh2002]

@toasted_toasts_ Idk why they prefer a punchy way of writing. Like everything has to be wrapped up like a revelation

[Nishant Sharma @nishantsh2002]

5/ I think the barrier to entry for game dev has always been high. From what Ive seen - you had to know game engines and programming before you could make anything. I had a crack at it myself :) making 2D games with basic sprites in Unity (inspired by Rain World ig). I find it really inspiring that ai will lower that barrier for everyone and creative people who have been gatekept can finally take a real stab at game dev!

[Nishant Sharma @nishantsh2002]

I think Tim’s right and the clearest picture of where this could lead is the OASIS from Ready Player One.

We can have so much more than current games. We can have digital worlds where anyone can go anywhere and build anything and I think that alone is incredibly inspiring and we should welcome ai to assist us building such worlds!

[Tim Sweeney @TimSweeneyEpic]

The games we've long wanted, which didn't exist because they'd cost several times more to make than they'd earn back, will soon become possible. As will the games developers wanted to build, but couldn't with the resources available. It'll be a new golden age of gaming.

[Nishant Sharma @nishantsh2002]

4/ In the book and the movie, Spielberg painted OASIS as a part of a dystopia. I dont think it has to be that way for us. If these worlds get immersive enough, people would want to spend real time in them.

And as manufacturing gets better and materials get cheaper, gear like VR headsets and omnidirectional treadmills will become affordable enough that it makes sense to own them (the way gaming consoles are today).

[Nishant Sharma @nishantsh2002]

@NC_Renic ngl those conjunctions make academic textmuch readable

[Nishant Sharma @nishantsh2002]

@LauraPGomez @ganstlr @apsarathchandar @HiRoKi11x @ChandarLab @Mila_Quebec Joint work with
@LauraPGomez_ , @ganstlr and my mentors @apsarathchandar and @_Hiroki11x. Thanks to
@ChandarLab and @Mila_Quebec for their support!

[Nishant Sharma @nishantsh2002]

New paper! 🧵

In communication-efficient distributed training, workers take several local steps, and an outer optimizer (often with Nesterov momentum) applies their averaged update to accelerate training.

We study when this is stable, and derive an exact criterion (in the mean-square sense).

[Nishant Sharma @nishantsh2002]

@LauraPGomez @ganstlr @apsarathchandar @HiRoKi11x @ChandarLab @Mila_Quebec Check it our here https://t.co/kRBhktTb9D!

[Nishant Sharma @nishantsh2002]

@benjamintherien Awesome!

[Nishant Sharma @nishantsh2002]

Beyond random quadratics, the same mean-square dynamics can be measured at checkpoints of real networks, via shared-minibatch perturbations (viewing training as a dynamical system)

We test the predictions on ResNets and Pythia LMs up to 1B parameters.

[Nishant Sharma @nishantsh2002]

Intuitively, averaging over more workers divides the noise in the outer update, so larger outer momentum becomes admissible.

But the number of workers needed for stability can grow exponentially with the number of local steps.