Twitxy Saved

Yao Lai

@chingyaolai

Assistant Professor @Stanford studying fluids, ML, and the physics of ice & climate.
Stanford, CA, USA icyphysics.stanford.edu Joined August 2020
532 Following 1.7K Followers 273 Posts

Latest posts 5

Yao Lai @chingyaolai
Prediction ≠ Insights. Despite the fact that AI agents excel at well-defined prediction tasks, they still struggle to distill insights from predictions like human scientists do, across a variety of domains. Check out this paper by @JiayiiGeng that asks this important question.
Jiayi Geng Verified account @JiayiiGeng Oct 2
AI scientist agents are great at optimizing and finding solutions through endless trial and error. But is that really what science is about?

As Terrance Tao eloquently put, the role of math and science should be more than that – they are "lighthouses" that guide and inspire exploration through understanding and insights. Can AI agents discover novel insights through experimentation?

To study these gaps, we worked with domain experts to introduce EurekaBench, a benchmark spanning 6 science domains that evaluates an agent’s ability to discover genuinely new scientific insights.
(1/n)
View quoted post on X
1 replies 4 reposts 54 likes 4.1K views View on X
Yao Lai @chingyaolai
Excited to see Terminal-Bench-Science finally released! It explores the types of tasks that can be solved relatively easily by scientific domain experts but remain challenging for AI agents. Kudos to @StevenDillmann for leading this effort!
#agenticAI #AI4Science
Steven Dillmann Verified account @StevenDillmann Aug 27
We're releasing Terminal-Bench-Science: a benchmark for evaluating AI agents on research workflows across scientific domains.

An ongoing Stanford-led community effort, built by the team behind Terminal-Bench together with scientific domain experts at research institutions worldwide. v0.1 has 70 tasks. Claude Opus 5 solves only ~30%.

1/n 👇
View quoted post on X
Yao Lai @chingyaolai
Just celebrated the graduation of my very first @Stanford PhD student, Niall. It feels even more joyful than receiving my own PhD!

Training the next generation of human scientists remains one of the most meaningful parts of academia- perhaps the most important in the age of AI.
6 replies 7 reposts 217 likes 24K views View on X
Yao Lai @chingyaolai
We're pleased to release DIFFICE-jax v1.0, the foundation of our Science paper and the first #DIFFerentiable #NeuralNetwork solver for #DataAssimilation of ICE shelves written in #JAX:
🔗Docs: diffice-jax.readthedocs.io/en/latest/inde…
📄Peer-reviewed by JOSS #OpenSource: joss.theoj.org/papers/10.2110…
Yao Lai @chingyaolai Mar 14, 2025
Finally published @ScienceMagazine:
Can AI help yield new insights from vast amounts of Earth data?
We use large-scale data and neural nets to find the constitutive laws of glacial ice, which differ from commonly assumed forms in conventional models.
science.org/doi/10.1126/sc…
View quoted post on X