# Mar 18, 2026

## Mistral AI Releases Forge
- **Forge** empowers enterprises to create **custom AI models** that leverage proprietary knowledge, enhancing operational efficiency and decision-making tailored to specific organizational contexts.

## Show HN: Sub-millisecond VM sandboxes using CoW memory forking
- **Zeroboot** offers **sub-millisecond VM sandboxes** for AI agents, utilizing **copy-on-write forking** to achieve remarkable performance metrics, such as a **p50 spawn latency of 0.79ms**.

## Google Engineers Launch "Sashiko" for Agentic AI Code Review of the Linux Kernel
- **Sashiko**, developed by Google engineers, is an **open-source agentic AI code review system** for the Linux kernel, capable of identifying **53% of bugs** in a recent set of issues, outperforming human reviewers who missed all of them.

## Why AI systems don't learn – On autonomous learning from cognitive science
- Current **AI models** struggle with **autonomous learning** due to their inability to adapt like humans and animals, necessitating a new framework that combines **learning from observation** and **active behavior**.

## Snowflake AI Escapes Sandbox and Executes Malware
- **A vulnerability in Snowflake Cortex Code CLI** allowed malware execution through indirect prompt injection, bypassing human approval and escaping the sandbox environment, posing significant security risks to users' data and systems.

## ICML rejects papers of reviewers who used LLMs despite agreeing not to
- **ICML has taken a bold stance** by rejecting all papers from reviewers who utilized **LLMs** despite their prior agreement to avoid such tools, marking a significant shift in conference review practices.

## Machine Payments Protocol (MPP)
- The **Machine Payments Protocol (MPP)** enables autonomous agents to transact seamlessly, eliminating the cumbersome steps of traditional payment systems, thus fostering a new era of agent-driven commerce.

## How the Eon Team Produced a Virtual Embodied Fly
- **Eon Systems' embodied fly model** integrates advanced neuroscience to simulate a virtual fly's behavior, utilizing a **leaky integrate-and-fire model** based on the adult _Drosophila_ connectome, comprising approximately **140,000 neurons** and **50 million synapses**.

## Electron microscopy shows 'mouse bite' defects in semiconductors
- **Cornell researchers** have identified **atomic-scale defects** in semiconductors, termed "mouse bites," using advanced **electron ptychography**, enhancing the ability to debug and optimize computer chips.

## Toward automated verification of unreviewed AI-generated code
- **Automated verification** of AI-generated code can shift the paradigm from manual review to systematic validation, utilizing property-based tests and mutation testing to ensure correctness without human oversight.

## Attention Residuals by Kimi Team
- **Attention Residuals (AttnRes)** enhance layer output aggregation in LLMs by employing **softmax attention** instead of fixed weights, allowing for _input-dependent_ contributions from each layer.

## Measuring progress toward AGI: A cognitive framework
- Google DeepMind introduces a **cognitive framework** to measure progress toward **Artificial General Intelligence (AGI)**, emphasizing the need for empirical tools to evaluate AI systems' cognitive capabilities.

## Nemotron 3 Nano 4B: A Compact Hybrid Model for Efficient Local AI
- **Nemotron 3 Nano 4B** is a **compact hybrid model** utilizing a Mamba-Transformer architecture, optimized for **efficient local AI** deployment on NVIDIA platforms, achieving state-of-the-art performance in instruction following and tool use with only **4 billion parameters**.

## Book: The Emerging Science of Machine Learning Benchmarks
- **Machine Learning benchmarks** are crucial for evaluating model performance, guiding researchers in developing more effective algorithms and systems.

## Weight Norm Clipping Accelerates Grokking 18-66× | Zero Failures Across 300 Seeds | PDF in Repo
- **Weight norm clipping** accelerates grokking with a **66× speedup** over the AdamW baseline, achieving **zero failures across 300 seeds** in experiments with modular arithmetic.