Gowthami
@gowthami_s
Followers
10K
Following
14K
Media
374
Statuses
3K
Multimodal research | Past - UMD, MetaAI, Amazon, IIT Madras | Rants, Memes my own.
Mountain View, CA
Joined April 2015
Excited to announce our MIT Press book “Neuroevolution: Harnessing Creativity in AI Agent Design” by Sebastian Risi (@risi1979), Yujin Tang (@yujin_tang), Risto Miikkulainen, and myself. We explore decades of work on evolving intelligent agents and shows how neuroevolution can
16
183
811
Two recent papers ( https://t.co/9yWbOj5sDF,
https://t.co/WID9ZW0kNQ) suggest that predicting x (clean) works much better than predicting eps or v (noisy) in high dimensions. Natural signals like images live on a low-dimensional manifold. Noise takes you off the manifold! (1/3)
arxiv.org
Today's denoising diffusion models do not "denoise" in the classical sense, i.e., they do not directly predict clean images. Rather, the neural networks predict noise or a noised quantity. In this...
17
77
550
Trying to decide what to do on the first day of #NeurIPS2025? Check out my, @ziqiao_ma, and @xiangyue96's tutorial, "The Science of Benchmarking: What's Measured, What's Missing, What's Next" on December 2 from 1:30 to 4:00pm. What will we cover? 1/3
1
28
194
Excellent paper with a simple story. They show that diffusion models are better when they output the pixel-prediction instead of noise/v prediction. The benefit of epsilon/v is from the loss function (probably due to variance reduction), so you predict x, but use v-loss.
Huge! @TianhongLi6 & Kaiming He (inventor of ResNet) just Introduced JiT (Just image Transformers)! JiTs are simple large-patch Transformers that operate on raw pixels, no tokenizer, pre-training, or extra losses needed. By predicting clean data on the natural-data manifold,
6
39
459
🥹this paper is really dear to my heart not just cause of the work, but of the team and how we came together! Renfei (right), first author, is applying to grad school this year, she is extremely brilliant, you should hire her (well I am going to as well so competition is on!)
9. RL Enhances Knowledge Navigation Researchers show that RL-enhanced models outperform base models by 24pp on hierarchical knowledge retrieval tasks by improving navigation of existing knowledge structures rather than acquiring new facts. https://t.co/E5HrlsTTce
5
9
234
Diffusion LMs are more data efficient than Autoregressive (AR) LMs, and the difference is insane. Due to its training objective, DLMs basically have built-in Monte Carlo augmentation. So when unique data is limited, DLMs would always surpass AR models, and harder to overfit!
8
52
369
A bit of context - In India, if you don't fit the mold—say, you don't have a certain degree or a high GPA—it's incredibly hard to break in. If you tell someone, "Please give me one chance, I’ll learn things and do well," you are far more likely to get that chance in the US.
@rao2z @iitmadras I studied mechanical engineering undergrad. I got into programming/ML when I did a startup in India. After the startup I applied to multiple programming/ML sort of roles in India, even small startups rejected me, cuz they wanted someone with experience rather than someone who is
6
3
78
Growing up in small town in India, you can see greatness from far, someone reached moon, nasa put a rover on mars… you can see the greatness but you are not sure you can be more…. We moved to US as an experiment… masters became PhD… and then my partner joined Optimus…
29
51
1K
The value of fast iteration in AI is overrated. The best results are obtained by knowing the right things to do and doing each thing with neurotic precision and attention to detail.
26
34
433
Accidentally said model instead of policy and they kicked me out of SF.
Accidentally used DPO instead of GRPO and they kicked me out of SF.
0
1
43
Abhinav is a solid researcher and a great PhD advisor. Work with him if you wanna get into GPUMode! :)
A large number of PhD students in my group have graduated or will be graduating by Spring, so I am recruiting several PhD students for the next admission cycle (Fall 2026). If you want to work with us, apply by Dec 5 and drop me a short email. Please repost/share widely. #HPC #AI
3
12
126
Reposting this evergreen meme of mine in honor of ICLR reviews
0
29
650
Well the number of future acceptances are a function of how many LLMs you have access to and how good you are at prompting I guess. LLM written papers reviewed by LLM reviewers. What’s the point of conferences even! 🤦♀️🤷♀️
This LLM-generated paper was submitted at least 4 times to ICLR 2026 with different titles, each with slight variations in content, but very similar core claims and (incorrect) proofs. 1/n
3
0
10
We disrupted a highly sophisticated AI-led espionage campaign. The attack targeted large tech companies, financial institutions, chemical manufacturing companies, and government agencies. We assess with high confidence that the threat actor was a Chinese state-sponsored group.
1K
3K
22K
I don’t know why this took 6 days to reach me - totally apt song describing the situation in Bay Area (or world in general). 😂😂
1
0
7