● ResearchOct 29, 2026 · 03:118 sources
Cartesia's New AI Architecture Cuts LLM Training Compute and Data Needs by 10x
From marketbrief.now
AI;DR
Cartesia has released an updated version of its Hierarchical Networks (H-Nets) architecture that allows large language models to learn from raw bytes, reportedly reducing training compute and data needs by 10x while matching the performance of a model trained on 11T tokens. The company also reported efficiency gains of 3-5x over existing methods in controlled scaling experiments.
Read originalSign in to suggest
Sources
+ 6 more
- SUPPORT@tri_dao“H-Net learns to dynamically tokenize variable names in coding data as separate tokens, and it scales much better on code compared to BPE”x.com
- SUPPORT@xennygrimmato_“H-Nets perform the chunking dynamically directly inside the model.”x.com
- SUPPORT@willdepue“as karpathy said, the ugliest part of the transformer is the tokenizer”x.com
- SUPPORT@realdanfu“More H-net work, very cool from @_albertgu and @cartesia research!”x.com
- SUPPORT@teortaxestex“H-Nets didn't work last year”x.com
- SUPPORT@xennygrimmato_“Btw here's the H-Net preprint from July 2025”x.com
#llmMarkdownScore 0.13.3k tokens
Content history (1)
- 2026-10-28 · Summary · vi · Replaced stored textCác nhà nghiên cứu Cartesia đã ra mắt phiên bản cập nhật của Hierarchical Networks, hay c…Các nhà nghiên cứu Cartesia đã ra mắt phiên bản cập nhật của Hierarchical Networks, hay c…