Twitxy Saved

Zihan Yan

Verified account This account is verified. Learn more

@physicscat0x7d

AI Scientist & Theoretical Physicist
MTS @ Stealth
Prev PhD/MA/MMath/BA @Cambridge_Uni
Cauchy slice z-yan.dev Joined October 2024
242 Following 861 Followers 192 Posts

Latest posts 5

Zihan Yan Verified account @physicscat0x7d
Great work by my colleagues! An inspiring direction to crack the fundamental nature of reasoning.

Congratulations @Ben_Hoov @HughBlayney @jcrwhittington and team @prefrontal_ai
Ben Hoover @Ben_Hoov 4d
Big reasoning speedups are here. 1,000× faster training and 280× less compute than any looped model. All done using a fully trainable energy from the 80s

Keep reading to hear how energy is the right paradigm for reasoning

With @HughBlayney @jcrwhittington @prefrontal_ai

🧵/10
View quoted post on X
Zihan Yan Verified account @physicscat0x7d
Super excited to see the release of LSA - great work by my colleagues at @prefrontal_ai. Looking forward to empowering AI scientists with it to reason across entire fields of literature and multi-million-token lab trajectories in a single pass.
James Whittington Verified account @jcrwhittington Nov 2
Today we are happy to share one of our first open research releases: Latent Sparse Attention (LSA), a step towards subquadratic attention that matches accuracy of full attention at ultra-long context.

Sparse attention attends to only a few tokens per query, but it has to find the right ones. At 10M tokens, that search takes 98% of the attention layer's time. LSA makes sparse attention up to 22.7× faster and full attention 72.4x faster.

With @kubicek_ales @caciolai @PengLiangzu @SukritSumant @Ben_Hoov @FraserGreenlee @RMBattleday @radim_urban @thoefler @jcrwhittington
View quoted post on X
Zihan Yan Verified account @physicscat0x7d
很开心和大家分享我从理论物理转AI之后参与的第一个项目:DiG-bench!

网址:digbench.ai
技术报告:arxiv.org/abs/2608.12593

我们设计了70个基于文本的游戏世界来测试大语言模型的发现能力:这些前沿模型是否可以通过主动探索和实验来发现游戏的规则和获胜目标,并且将这些发现转化为行动最终通关呢?尽管过去的几个月里大模型日新月异、突飞猛进,目前最好的模型还是会在一些人类看来很简单的游戏世界里栽跟头。我们希望以此为原型来检验模型在“发现”这方面的短板,更好地推进AI的科学发现能力,造出更好的scientific discovery agent!
digbench.ai dig.bench DiG-bench (Discovery in Games): a benchmark of scientific discovery — 70 interactive games with undisclosed rules, played by humans and frontier AI models through the same interface.