reduce sizes of gif

Author: YWolfeee

_pages/about.md

@@ -23,7 +23,7 @@ Feel free to reach out if you are interested in my research or simply want to ch
 <h2 id="news">News</h2>
 - (Jan. 2026) We release an interesting [study](https://linhaowei1.github.io/scaling_law_discovery/) on whether AI can discover its own scaling laws better than humans. Check it out!
 - (Jan. 2026) Check out [InfoTok](https://arxiv.org/abs/2512.16975), our adaptive information-theoretic tokenizer that achieves 40% higher compression without information loss and represents videos via coarse-grained to fine-grained sequences.
-- (Dec. 2025) We release [DDRL](https://research.nvidia.com/labs/dir/ddrl/), a data-regularized RL algorithm for diffusion! I use DDRL to post-train the NVIDIA [Cosmos-Predict2.5](https://research.nvidia.com/labs/dir/cosmos-predict2.5/) models (2B/14B) with ~2K H100 GPUs. Check our [post](https://x.com/haotian_yeee/status/1997138677529825452?s=20)!
+- (Dec. 2025) We release [DDRL](https://research.nvidia.com/labs/dir/ddrl/), the RL algorithm powering NVIDIA's [Cosmos-Predict2.5](https://research.nvidia.com/labs/dir/cosmos-predict2.5/) video generative models (2B/14B)! An amazing experience scaling the post-training of models on ~2,000 H100 GPUs. Check our [post](https://x.com/haotian_yeee/status/1997138677529825452?s=20) and [models](https://github.com/nvidia-cosmos/cosmos-predict2.5)!
 - (Dec. 2025) [Data Attribution for RL](https://arxiv.org/abs/2505.19281) was accepted by NeurIPS 2025 (<font color="#DC143C">Oral</font>). See you in San Diego!
 <!-- - (Nov. 2025) We released NVIDIA [Cosmos-Predict2.5](https://research.nvidia.com/labs/dir/cosmos-predict2.5/), the Cosmos World Foundation Models specialized for video generation. I am responsible for designing the RL algorithm and performing large-scale post-training (1K+ GPUs) for the release. -->  
 <!-- - (Oct. 2025) [Diffusion Inference-Time Acceleration](https://arxiv.org/abs/2507.15260) was accepted by ICCV 2025. See you in Hawaii!  -->