Dheeraj Varghese

PhD candidate @UvA_Amsterdam

[Dheeraj Varghese @dhevarghese]

@zzigakovacic @OpenAI Soon! 👀

[Dheeraj Varghese @dhevarghese]

Ever wondered whether AI agents could play thousands of video games on a supercomputer? We did. 🎮

Introducing PlaySuite: 5,734 games, 14 open models, and a large-scale benchmark for interactive visual intelligence.

Here's what we built, and what we learned. 🧵👇

[Dheeraj Varghese @dhevarghese]

A huge shoutout to the wonderful team behind PlaySuite! 🎮

It was a pleasure working with @AVettoruzzo, @iWSimoncini, Michelle Acevedo Callejas, @mmderakhshani, Kristof Meding, @joavanschoren, and @cgmsnoek.

Thanks for making this possible! 🙌

[Dheeraj Varghese @dhevarghese]

Nearly a decade after @OpenAI Universe, PlaySuite revisits the vision of evaluating general-purpose agents across thousands of interactive worlds.

Can your agent play games it's never seen before? 🎮

📄 https://t.co/4hZfyWsB96

[Dheeraj Varghese @dhevarghese]

@pmpcurvo Very interesting!!

[Pedro @pmpcurvo]

♾️ LiFT: Loop Flow Transformers 🧵👇

Diffusion models usually spend more inference compute by taking more sampling steps. What if we could also spend compute inside each step?

LiFT turns part of a DiT into a recurrent core, letting us scale depth at inference with only a small architectural change. On ImageNet, it beats dense DiT-XL/2 with 3.34 lower FID, while using 60% fewer parameters, 32% fewer training FLOPs, and 52% fewer inference FLOPs.

Jointly led with Mohammad Mahdi Derakhshani 👑(@mmderakhshani)

And amazing supervisors Cees Snoek (@cgmsnoek), Jan-Willem van de Meent (@jwvdm) and Gertjan J. Burghouts (@gjburghouts)

[Dheeraj Varghese @dhevarghese]

Excited to be a part of this! ✨

[Cohere Labs @Cohere_Labs]

3 days. Worldwide. Inspiring & starting new research collaborations. Introducing the Connect conference. 🖇️

Join for incredible speakers, including @1vnzh @jpineau1 @mziizm & @ShayneRedford + >20 researchers discussing how collaboration and open science are driving progress. 🚀

[Dheeraj Varghese @dhevarghese]

Through NeoBabel, I’ve experienced how large-scale distributed training works in practice. @mmderakhshani's walkthrough distills those lessons, illustrating the five key parallelism strategies. A great read👇

[Mohammad Mahdi (Medi) Derakhshani @mmderakhshani]

🔗 Inside 3D Parallelism: Scaling Transformers in PyTorch (A Code-Study Walkthrough):
https://t.co/UfEN0HGNYU

[Yulu Qin @yulu_qin]

Does vision training change how language is represented and used in meaningful ways?🤔 The answer is a nuanced yes! Comparing VLM-LM minimal pairs, we find that while the taxonomic organization of the lexicon is similar, VLMs are better at _deploying_ this knowledge. [1/9]

[Cohere Labs @Cohere_Labs]

🖼️ What does it take to build a multilingual image generation model that is both efficient and competitive?

🌍 In collaboration with @UvA_Amsterdam, we introduce NeoBabel, a 2B multilingual text-to-image model trained natively in 6 languages.

[Mohammad Mahdi (Medi) Derakhshani @mmderakhshani]

T1/11
Text-to-image AI is amazing, but what if you don't speak English? Most are English-centric, creating barriers for billions.
We introduce NeoBabel, a new open model for visual generation that natively understands 6 languages, breaking down the walls of digital inequity. 🌍🗣️

[François Chollet @fchollet]

You can learn something interesting from pretty much anyone