Stream: community

Topic: Meeting "Can the Physical World Make AI More Trustworthy?"

Ben Lewis (Apr 30 2026 at 19:29):

I'm sending out a much belated invite to this satellite session about ML-actionable metdata, with special consideration of the spatial and temporal dimensions (aka physical world). Here is a bit more information on the session. Excited for the conference! Cheers, Ben

Slava Tykhonov (May 01 2026 at 22:18):

Ben Lewis said:

I'm sending out a much belated invite to this satellite session about ML-actionable metdata, with special consideration of the spatial and temporal dimensions (aka physical world). Here is a bit more information on the session. Excited for the conference! Cheers, Ben

Ben, I don't know about physical world but I was asking the question about responsibility in the age of AI https://www.linkedin.com/posts/vyacheslavtikhonov_lets-imagine-youre-asking-your-personal-activity-7455893630076092416--EOf

Slava Tykhonov (May 01 2026 at 22:19):

See you there!

Ben Lewis (May 05 2026 at 16:11):

Thank you, @Slava Tykhonov!

Here is a related question: how do we have responsibility without auditability and reproducibility?

That is a hard problem for LLMs, but potentially less so for spatially grounded inferencing systems.

Which brings me to a quick reminder: sign-ups for satellite meetings have been extended until Thursday.
**Can the Physical World Make AI More Trustworthy?
Dataverse, Metadata, and the Infrastructure for Grounded Inference**

Data describing the physical, spatial, and mapped world comes in many surprising forms and is part of many research workflows.

This discussion will focus on how Dataverse can better support spatial, temporal, observational, and event-based data through machine-actionable metadata, provenance, uncertainty, and related external workflows.

**I will also touch on a few concrete experimental directions:

• the scope of spatial data in Dataverse, including tabular files with latent spatial structure
• practical generation of GeoCroissant for different file types
• a lightweight external “shadow Dataverse” layer that allows real systems to test ML-ready metadata and MCP-based access before anything is pushed into core Dataverse**

My own examples come mainly from earth and social-science use cases, but I would be especially glad to have participation from astronomy and other observational domains, where uncertainty, provenance, spherical coordinates, and machine reuse may already be more mature.

The format will be informal and discussion-oriented, with brief framing remarks followed by open discussion on practical next steps, useful pilots, and collaborations.

Please sign up by Thursday if you would like to attend: blewis@cga.harvard.edu

Last updated: May 30 2026 at 09:11 UTC