Announcing:
*The Polymathic AI Initiative*
We are researching a new class of foundation models for scientific *data*, developing models that share scientific concepts across disciplines.
Site:
News:
Thread: 👇
We're thrilled to be at
@NeurIPSConf
for the first time since we formed a few months ago!
If you're there, we'd love to chat with you about our team and AI/ML research!
Reach out and be sure to check out our accepted papers👇
#NeurIPS2023
#AI4Science
and we are HIRING 🥳
Are you a student looking for an internship this summer or fall?
Want to build foundation models for science in NYC?
Join us at
@PolymathicAI
at
@FlatironInst
!
There's much preliminary research required to build a foundation model for scientific data. This week are releasing three papers on the fundamentals of this space:
(1) xVal: continuous tokenization for numbers
(2) Multiple Physics Pre-training
(3) AstroCLIP
5/n
Our aim with the Polymathic AI initiative is to accelerate the development of versatile foundation models tailored for machine learning tasks in *science*. A breakthrough in this field would mark a significant shift analogous to the one seen in the domains of vision and NLP.
1/n
1/ Did you know LLMs like ChatGPT struggle with number manipulation and data analysis? They make up answers and fail to generalize.
Today I am excited to share xVal, a continuous and efficient number encoding scheme with better generalization performance.
Check out our oral presentation of Multiple Physics Pretraining (MPP) at
#AI4Science
workshop!
MPP is a group of techniques for pretraining on a diverse set of time-dependent physics to learn models that can be fine-tuned for multiple problems.
Very excited to present Multiple Physics Pretraining!
MPP is a group of techniques for pretraining on a diverse set of time-dependent physics to learn models that can be fine-tuned for multiple problems.
Blog:
Paper:
We are hiring a software engineer or software-focused researcher to embark with us on this journey of changing how science is done with large foundation models!
Reach out if you are excited!
A large number of our team is at NOLA this week!
Find us :
Very excited to present Multiple Physics Pretraining!
MPP is a group of techniques for pretraining on a diverse set of time-dependent physics to learn models that can be fine-tuned for multiple problems.
Blog:
Paper:
And got problems with numerics with LLMs?
Check out our xVal presentation at
#AI4Science
Workshop! where we discuss how to get efficient numeric encoding that generalizes well!
Blog:
1/ Did you know LLMs like ChatGPT struggle with number manipulation and data analysis? They make up answers and fail to generalize.
Today I am excited to share xVal, a continuous and efficient number encoding scheme with better generalization performance.