Nick Wagner
@science-nick.bsky.social
AAAS Science and Tech Policy Fellow at NSF. Views are mine alone.
Are these technical solutions inherently dual use for offense as well as defense? My understanding is Anduril uses Lattice to glue together all their systems. Not just the ones involved with defending stationary positions. You would probably need a different dataset for fine-tuning on targets though
December 5, 2024 at 5:28 AM
Are these technical solutions inherently dual use for offense as well as defense? My understanding is Anduril uses Lattice to glue together all their systems. Not just the ones involved with defending stationary positions. You would probably need a different dataset for fine-tuning on targets though
I should say 11B parameters for the smallest models I know. For high accuracy requirements (like minimizing false negatives on drone detections), you probably need even bigger models.
December 5, 2024 at 5:16 AM
I should say 11B parameters for the smallest models I know. For high accuracy requirements (like minimizing false negatives on drone detections), you probably need even bigger models.
The multimodal models to my knowledge are still quite memory-intensive (~11B parameters that are not easily quantizable). While you can defend a base with a fast internet connection or on-prem cluster, a remote base or mobile platform might struggle to run these models.
December 5, 2024 at 5:14 AM
The multimodal models to my knowledge are still quite memory-intensive (~11B parameters that are not easily quantizable). While you can defend a base with a fast internet connection or on-prem cluster, a remote base or mobile platform might struggle to run these models.