Mohammad Saffar (@msaffar3) / X

Mohammad Saffar

189 posts

Mohammad Saffar

@msaffar3

Research Scientist @googledeepmind, Gemini multi modal | past: @reveimage, Google brain

Mountain View, CA

Joined July 2016

Pinned
Mohammad Saffar
@msaffar3
Sep 3, 2025
After a fantastic year at Reve AI, I’ve rejoined Google DeepMind to continue working on VEO. I was deeply involved in its early days, but the rapid progress from VEO 1 to VEO 3 in just one year has truly amazed me. It’s a testament to what can happen when you combine compute,
25K
Mohammad Saffar
@msaffar3
Mar 24, 2025
Typography in generative models has been a hard problem to solve. Reve Image 1 does not only generate long text, but also it is very good at stylized typography. You can generate designs like this, it follows the prompt faithfully.
Taesung Park
@Taesung
Mar 24, 2025
Excited to come out of stealth at @reveimage! Today's text-to-image/video models, in contrast to LLMs, lack logic. Images seem plausible initially but fall apart under scrutiny: painting techniques don't match, props don't carry meaning, and compositions lack intention. (1/4)
14K
Mohammad Saffar
@msaffar3
Mar 24, 2025
Creative generative media has been my passion since my early days at Google Brain Research. I am beyond excited to finally share what we have been building for the past few months! The team is very small but the most creative and scientifically rigorous 😍 This is just the
Reve
@reve
Mar 24, 2025
Halfmoon is Reve Image — and it’s the best image model in the world 🥇 (🔊)
00:00
6.3K
Mohammad Saffar
@msaffar3
Jun 3, 2025
Replying to @eliebakouch
This is actually a score on not sharing "capabilities" research so they got a 0 🤷🏻‍♂️
1.6K
Mohammad Saffar
@msaffar3
May 14, 2024
This was the most fun I had working on a project so far!!
Google DeepMind
@GoogleDeepMind
May 14, 2024
Introducing Veo: our most capable generative video model. 🎥 It can create high-quality, 1080p clips that can go beyond 60 seconds. From photorealism to surrealism and animation, it can tackle a range of cinematic styles. 🧵 #GoogleIO
00:00
1.6K
Mohammad Saffar
@msaffar3
Jul 8, 2025
I think it is time to revisit scaling laws and incorporate FLOPs as a separate variable. Parameter and token counts alone do not give a full view as we can change FLOPs independently. Legendary approaches like 2-simplicial attn paper proposed by @Happylemon56775 and the
2K
Mohammad Saffar
@msaffar3
Jun 8, 2023
Replying to @upstuckGPT and @YiTayML
I guess you might want to look up Yi's Google scholar.
451
Mohammad Saffar
@msaffar3
May 21, 2025
Gemini diffusion is exciting! Letting models revise their answer is a powerful paradigm. Autoregressive models try to do this with CoT but diffusion makes this the main training objective. The main missing piece could be training efficiency. Can we make a single training step of
870
Mohammad Saffar
@msaffar3
Jun 21, 2024
This is cool! While folks are rightfully excited about int8 training, I want to highlight the importance of local attention. @ashVaswani @nikiparmar09 @NoamShazeer et al. where the first ones introducing it in the image transformer paper. It is hugely overlooked unfortunately.
Noam Shazeer
@NoamShazeer
Jun 20, 2024
Character AI is serving 20,000 QPS. Here are the technologies we use to serve hyper-efficiently. [research.character.ai/optimizing-inf… ]
1K
Mohammad Saffar
@msaffar3
Jul 4, 2025
Really nice work from my old friends! A powerful paradigm to escape the limits of token counts in training data is to increase token efficiency on the architecture side.
Aurko Roy
@aurko79
Jul 4, 2025
Excited to share what I worked on during my time at Meta. - We introduce a Triton-accelerated Transformer with *2-simplicial attention*—a tri-linear generalization of dot-product attention - We show how to adapt RoPE to tri-linear forms - We show 2-simplicial attention scales
688
Mohammad Saffar
@msaffar3
Dec 13, 2023
This is very exciting! Congrats @ashVaswani and @nikiparmar09. Looking forward to see the advancements by essential.ai 🙌🚀
Ashish Vaswani
@ashVaswani
Dec 12, 2023
I'm thrilled to announce our company, @essential_ai . We believe that breakthroughs in AI will unlock the most profound tools for thought, advancing humanity's collective knowledge and capability. essential.ai
GIF
516
Mohammad Saffar
@msaffar3
Jun 1, 2025
Replying to @_arohan_
431
Mohammad Saffar
@msaffar3
Apr 27, 2022
Replying to @ashVaswani
Congratulations Ashish, @nikiparmar09 and @jluan and the team! This is very exciting! Cannot wait to see the impact soon!
Mohammad Saffar
@msaffar3
Nov 29, 2023
Replying to @_arohan_ and @pika_labs
Generating articulate motion is hard, finetuning an image model for videos needs a lot of effort, maybe people should treat the video modality a little more natively.
650