24 October 2026

AI · Security

Claude Security now scans code continuously

Anthropic's Claude Security is now in public beta for Claude Enterprise. It scans code continuously as it changes, tracks fixes, and reopens a finding if the bug comes back. Anthropic says thousands of organizations have used it since April.

• Claude Security now scans continuously and tracks fixes Anthropic · 21 Oct 2026 · Primary

AI · Robotics

AI agents teach a robot to walk in simulation

NVIDIA shows AI agents training the small open-source Microduck robot to walk. The Hermes agent, running a Nemotron model, planned the training. The walking policy trained in Isaac Lab across 4,096 parallel simulated environments. So far it was validated in simulation only.

• Using AI Agents to Teach Robots Like Microduck to Walk in Simulation NVIDIA Technical Blog · Oct 2026 · Primary

AI · Architectures

KronLSTM: a sharper memory update

Researchers at JKU Linz and NXAI propose KronLSTM, a new member of the xLSTM family. Its Kronecker delta rule updates memory along a single direction, so new information interferes less with what is stored. On long contexts it is competitive with, and often better than, Gated DeltaNet.

• Kronecker Delta Rule for Long Context Linear Attention arXiv · 19 Oct 2026 · Primary · Hartl et al. (2026); arXiv 2610.26834, 19 Oct 2026

AI · Language models

Byte models break the token ceiling

Meta FAIR and University of Washington researchers revisit language models that read raw bytes instead of tokens. Token models plateau, while byte models start worse but have a higher ceiling. Distilled from a token model, a one-billion-parameter byte model matches it with one sixth of the data.

• Breaking the Token Ceiling: Distilling Smaller, Stronger Byte Models arXiv · 11 Sept 2026 · Primary · Marathe et al. (2026); arXiv 2609.12303, 11 Sep 2026

Idea · Machine learning

Knowledge distillation

Knowledge distillation. A big teacher model trains a small student, not only on the right answers but on its full probabilities. Shown a handwritten two, the teacher might say: almost surely a two, a tiny chance of a three, even less of a seven. Those soft targets reveal which mistakes are reasonable, information a plain label throws away. That is how this week's byte model learns from a token model.

Hinton, Vinyals & Dean 2015 · arXiv:1503.02531

• Distilling the Knowledge in a Neural Network arXiv:1503.02531 · Mar 2015 · Primary · Hinton, Vinyals & Dean (2015)

Idea · Architectures

Memory that corrects itself: the delta rule

Linear attention can be read as a memory matrix. Each new key and value is simply added on top, so over a long context the entries pile up and blur. The delta rule first reads what the memory already returns for that key, then writes only the error. Like crossing out an old phone number instead of writing a second one beside it. KronLSTM refines exactly this update.

Schlag, Irie & Schmidhuber 2021 · arXiv:2102.11174

• Linear Transformers Are Secretly Fast Weight Programmers arXiv:2102.11174 · Feb 2021 · Primary · Schlag, Irie & Schmidhuber (2021)

Idea · Robotics

Domain randomization

A simulator is never exactly the real world. Instead, make it varied: random textures, lights, masses and friction in every run. If a model copes with all of them, reality may look like just another variation. In 2017, a detector trained only on simulated images located real objects to within one and a half centimetres. Microduck's training uses the same trick.

Tobin et al. 2017 · arXiv:1703.06907

• Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World arXiv:1703.06907 · Mar 2017 · Primary · Tobin et al. (2017)