Ziyu Wang
418 posts
@ziyuwang
Machine learning researcher. Co-founder / CTO of https://t.co/h85fQdqLrr
🚨New model drop into the Top 10! 🖼️ @MicrosoftAI just entered the Image Arena with MAI-Image-1. Community votes are already rolling in, and MAI-Image-1 has broken into the Top 10. It’s currently ranked #9, tied with Seedream 3! MAI-Image-1 is now live in Direct Chat for early access on LMArena.
Meet our third @MicrosoftAI model: MAI-Image-1 #9 on LMArena, striking an impressive balance of generation speed and quality Excited to keep refining + climbing the leaderboard from here! We're just getting started. microsoft.ai/news/introduci…
Here's my 5-hour conversation with @dylan522p and @natolambert on DeepSeek, China, OpenAI, NVIDIA, xAI, Google, Anthropic, Meta, Microsoft, TSMC, Stargate, megacluster buildouts, RL, reasoning, and a lot of other topics at the cutting edge of AI. This is was a mind-blowing, super-technical, and fun conversation. Yes, we discuss r1 and o3-mini, but more importantly we look into the future of technology, geopolitics, and humanity in a world that stands on the precipice of a global AI revolution. The first 4 hours are here on X (4 hours is current limit), and the full 5 hours are up everywhere else. Links in comment. Timestamps: 0:00 - Introduction 3:33 - DeepSeek-R1 and DeepSeek-V3 25:07 - Low cost of training 51:25 - DeepSeek compute cluster 58:57 - Export controls on GPUs to China 1:09:16 - AGI timeline 1:18:41 - China's manufacturing capacity 1:26:36 - Cold war with China 1:31:05 - TSMC and Taiwan 1:54:44 - Best GPUs for AI 2:09:36 - Why DeepSeek is so cheap 2:22:55 - Espionage 2:31:57 - Censorship 2:44:52 - Andrej Karpathy and magic of RL 2:55:23 - OpenAI o3-mini vs DeepSeek r1 3:14:31 - NVIDIA 3:18:58 - GPU smuggling 3:25:36 - DeepSeek training on OpenAI data 3:36:04 - AI megaclusters 4:11:26 - Who wins the race to AGI? 4:21:39 - AI agents 4:30:21 - Programming and AI 4:37:49 - Open source 4:47:01 - Stargate 4:54:30 - Future of AI
A General Framework for Inference-time Scaling and Steering of Diffusion Models Introduces Feynman-Kac steering, an inference-time steering framework for sampling diffusion models guided by a reward function. It generates multiple samples (particles) like best-of-n (importance sampling) approaches. Particles are evaluated at intermediate steps, where they are scored with functions called potentials. Potentials are defined using intermediate rewards and are selected such that promising particles are resampled and poor samples are terminated. "FK steering with just k = 4 particles outperforms fine-tuning on prompt fidelity and aesthetic quality, without making use of reward gradients." "FK steering smaller diffusion models outperforms larger models, and their fine-tuned versions, using less compute."
