Extropic unveils Z1T claiming 140x energy efficiency gains, but critics call it a sparse quantized transformer
burny_tech · x · 2026-09-06
Extropic introduced Z1T, its first transformer-like model family built for sparse probabilistic hardware (Z1), claiming up to 140x energy efficiency gains over GPUs and a new scaling law for sparse transformers. Critics immediately pushed back: one reply argued it's 'basically just a sparse quantized transformer' with no probabilistic sampling, so the p-bit noise adds nothing of value — questioning the headline efficiency claims.
More from Infra
- Cathie Wood claims Anthropic pays ~$50B per gigawatt while xAI's build costs sit in the high-20s — Kyrannio · 2026-09-06
- Mixing a 3090 With RTX 50 Cards for Local LLM Inference: The Performance Trade-offs — mrgreatheart · 2026-09-06
- Nvidia invests $3.5B in MediaTek as chipmaker joins NVLink Fusion ecosystem — Beth_Kindig · 2026-09-06
- Duplicating one Pinokio app and deduping saved 140GB on disk — cocktailpeanut · 2026-09-06
- AWS to build 50MW AI Zone in Saudi Arabia with HUMAIN, mixing Trainium and Nvidia GPUs — Beth_Kindig · 2026-09-06
- GPT-3 to GPT-6: OpenAI API prices dropped up to 6x in five years — BLUECOW009 · 2026-09-06