MarkTechPost

NVIDIA Shows 3 Ways to Speed Up MoE Inference on DGX Spark, Lifting Ling-3.0-flash Throughput 42.5%

NVIDIA shows 3 ways to speed up MoE inference on DGX Spark, lifting Ling-3.0-flash GSM8K decode throughput 42.5%.
GitHub

README.md

Geometry of confabulation and refusal in small language models: ConfabQA benchmark (784 items), linear probes scored against prompt-feature baselines, SAE decomposition, and refusal steering on ...
GitHub

Encoding a rotation: replica theory and out-of-distribution error of a linear encoder

How does a linear encoder learn the angular velocity of a rotating object from raw video, and can it extrapolate to velocities it has never seen? Replica theory, exact linear-algebra identities and ...
X

mccoy on X: "@GhostExecutes Even a 5 year old could decode this Imma start trading live , just to make them puke"

ENIGMA👁️ Time isn't linear, it's nested. 90 minutes split 1:2 over and over, with each circle holding a smaller version of the same move. Watch the playhead cross 9:30 → 11:00. 👇 ...
Tech Times

Graph Architecture Replaces Transformer Attention, Extrapolating to 4 Million Tokens in Linear Time

GTFormer proposes replacing Transformer self-attention with a hierarchical graph neural network, achieving zero-shot length ...
BIOENGINEER

AI and Omics Converge to Decode the Gene Regulatory Logic of the Brain

A new Brain Informatics review traces how AI methods, from co-expression statistics to causal graph neural networks and ...