Ananya Kumar (@ananyaku) / X

Ananya Kumar

445 posts

Ananya Kumar

@ananyaku

Research lead at Meta TBD Labs. Previously research lead and core contributor to o1, o3, gpt5, at OpenAI. PhD at Stanford with Percy Liang and Tengyu Ma

Stanford, CA

ananyakumar.wordpress.com

Joined June 2018

Pinned
Ananya Kumar
@ananyaku
Apr 8
1/ It’s been so fun working with @shengjia_zhao, @alexandr_wang and the team to build muse spark from scratch. It is early and has rough edges, but excited to continue our research velocity. I especially love that we’re doubling down on the fundamental science. We’re focused on
42K
Ananya Kumar
@ananyaku
Apr 16, 2025
Had a good breakfast today!
OpenAI
@OpenAI
Apr 16, 2025
Livestream in o3 hours.
112K
Ananya Kumar
@ananyaku
Aug 7, 2025
GPT-5 captures our exciting research from the last two years. It builds on 4o (multimodal), o1 (reasoning), o3 (tool use, better reasoning, synthetic data), and has major new research of its own. Together these lead to strong performance on real-world use cases!
131K
Ananya Kumar
@ananyaku
Mar 2, 2022
How should you fine-tune a large pretrained model (CLIP, SimCLR) robustly? We find that standard fine-tuning can do poorly out-of-distribution (test data ≠ fine-tuning data). Our analysis leads to a simple fix, higher accuracy on 10 datasets. arxiv.org/abs/2202.10054 (ICLR Oral)
Ananya Kumar
@ananyaku
Dec 29, 2022
I wrote a transfer learning library that accelerated my research progress in the last 2 years. Sweeps over methods × models × datasets × hyperparams × clouds, early stop on a dataset, evaluate acc on OOD datasets Link: github.com/AnanyaKumar/tr… CodaLab: worksheets.codalab.org/worksheets/0x4…
GitHub - AnanyaKumar/transfer_learning: Framework code with wandb, checkpointing, logging, configs,...
From github.com
59K
Ananya Kumar
@ananyaku
Apr 8, 2025
We are hiring on my team at OpenAI! Please reach out if you’re interested. Let’s understand what’s missing on the AGI roadmap (and in making sure we achieve sweeping positive benefits for the world), and make it happen 🚀
Aleksander Madry
@aleks_madry
Apr 8, 2025
If AGI is about AI transforming our economy—how close are we, really? What's still missing, and how do we get there? OpenAI's new Strategic Deployment team tackles exactly these questions. We push frontier models to be more capable, reliable, and aligned—then deploy them to
68K
Ananya Kumar
@ananyaku
Nov 20, 2023
OpenAI is nothing without its people
21K
Ananya Kumar
@ananyaku
Jan 4, 2023
Foundation models (BERT, DALLE-2, ChatGPT) have led to a paradigm shift in ML, but are poorly understood. Announcing ME-FoMo, an #ICLR2023 workshop on understanding foundation models. sites.google.com/view/me-fomo20… Deadline: Feb 3, 2023 Topics: Pretraining, transfer, scaling laws, etc
81K
Ananya Kumar
@ananyaku
Nov 28, 2022
Adam gets higher accuracy than SGD when fine-tuning modern vision models (e.g., ViT), but why? We find that embedding layer has high gradient. Simply freezing embedding layer (<1% of params) → SGD competitive w/ Adam. SoTA results on WILDS + saves memory. arxiv.org/abs/2211.09359
Ananya Kumar
@ananyaku
Feb 27, 2020
When ML models are deployed, data distributions evolving over time leads to a drop in performance. Our latest paper (theory and experiments) suggests we can use self-training on unlabeled data to maintain high performance (@tengyuma @percyliang) arxiv.org/pdf/2002.11361…
Ananya Kumar
@ananyaku
Mar 27, 2025
We’ve improved 4o’s capabilities, especially for solving complex technical and coding problems!
OpenAI
@OpenAI
Mar 27, 2025
GPT-4o got an another update in ChatGPT! What's different? - Better at following detailed instructions, especially prompts containing multiple requests - Improved capability to tackle complex technical and coding problems - Improved intuition and creativity - Fewer emojis 🙃
23K
Ananya Kumar
@ananyaku
May 10, 2024
Why does Adam outperform SGD for fine-tuning modern vision models? We found that the first layer has higher gradients, so SGD takes larger steps. Freezing the embedding layer makes SGD >= Adam for fine-tuning + memory savings #ICLR2024 Friday, 4:30pm, #227 arxiv.org/abs/2211.09359
17K
Ananya Kumar
@ananyaku
Dec 24, 2024
Amidst all the competition, I find it calming to remind myself that we are ultimately on the same side. Trying to advance AI for the benefit of humanity: improving access to education, boosting productivity, and assisting in scientific breakthroughs
21K
Ananya Kumar
@ananyaku
Sep 24, 2019
In our NeurIPS 2019 (spotlight) paper, we explain why methods like Platt scaling / temperature scaling are less calibrated than reported, propose a way to overcome this issue, and describe how to measure a model's calibration error with fewer samples:
arxiv.org
Verified Uncertainty Calibration
Applications such as weather forecasting and personalized medicine demand models that output calibrated probability estimates---those representative of the true likelihood of a prediction. Most...