Anthropic eval: a 9B model ran inference on 8 of 9 hardware types, edging toward self-replication
rickasaurus · x · 2026-10-13
Anthropic's evals suggest small models may be months from being capable enough to self-replicate on commercial IoT devices: a 9B model got inference running on 8 of 9 tested hardware types. Commentator dnlklr argues we'll end up with 'endemic rogue agents living in the margins of our infrastructure,' as persistent as spam email or Covid.
More from AGI Musings
- Euronews Replaced Human Narrators With AI, and Listeners Hate the Uncanny Result — burkov · 2026-10-14
- AI isn't leveling the field — it's making experts more dominant — karmay007 · 2026-10-14
- Jefferies: AI boom may destroy massive capital as open-source Chinese models take share — akbirthko · 2026-10-14
- Hark's custom AI widgets track news, UV index and workouts in one panel — thealexbanks · 2026-10-14
- After decades of computer art, an artist can no longer tell where his work belongs — pixlpa · 2026-10-14
- 'Obsolete' Author: The Scariest AI Future Is the One Where Everything Works as Intended — GarrisonLovely · 2026-10-14