Observed arrival · 2026-09-09
Geocentric’s AI Models Come With Their Mistakes Attached
An AI safety lab publishes language models, evaluation results, training research, and the infrastructure used to serve them.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
Field notes
Arc 120M is served through Geocentric’s cloud and is accompanied by a model card containing 24 unedited responses, including failed outputs. The lab’s systems work uses controlled Apple M4 comparisons: one reported change moves generation from 74.26 to 88.27 tokens per second, while another reduces optimizer-state memory by 65.4% in an 11.2M-parameter test. Additional work is organized under named systems including EPICYCLE and PARALLAX.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue