AI Daily — April 21, 2026
2026-04-21
rss.arxiv.org
BASIS: Decoupling Activation Memory from Batch and Sequence Dimensions in Backprop
BASIS (Balanced Activation Sketching with Invariant Scalars) is a new backpropagation algorithm that breaks the O(L×BN) activation memory bottleneck by fully decoupling memory from batch and sequence dimensions while still propagating exact error signals (dX). This addresses a longstanding constraint that limits scaling of deep networks with long contexts or large batches. If the claims hold up, this could have meaningful implications for training large models under memory-constrained regimes.
rss.arxiv.org
ETC: Scalable Concept Erasure for Text-to-Image Diffusion Models at Thousands of Concepts
Erasing Thousands of Concepts (ETC) presents a framework for removing undesirable or copyrighted content from text-to-image diffusion models at a scale previously impractical — existing methods top out around a few hundred concepts. The approach models low-rank concept distributions via a Student's t-distribution Mixture Model (tMM) to enable precise affine-based erasure without degrading generation quality. This is a meaningful advance for safety and compliance tooling around deployed generative models.
jack-clark.net
Import AI 454: Automating Alignment Research, Chinese Model Safety Study, Huawei HiFloat4
Jack Clark's latest Import AI covers three notable threads: progress on automating alignment research, a safety evaluation of a Chinese frontier model, and Huawei's HiFloat4 low-precision training format which reportedly outperforms MXFP4 on Ascend chips. The HiFloat4 result is a concrete signal that export controls are pushing Chinese hardware/software co-design in a distinct direction from Western ML infrastructure. The alignment automation angle is worth watching as labs explore using AI to accelerate safety research itself.