# Mar 19, 2026

## Waymo Safety Impact
- **Waymo's safety data reveals a significant reduction in crash rates**, with a **92% decrease in serious injury or worse crashes** compared to human drivers, showcasing the effectiveness of autonomous driving technology in enhancing road safety.

## Nvidia greenboost: transparently extend GPU VRAM using system RAM/NVMe
- **nvidia_greenboost** is a project with **34 commits** and **1 branch**, showcasing active development in optimizing NVIDIA GPU performance.

## Show HN: Duplicate 3 layers in a 24B LLM, logical deduction .22→.76. No training
- **Duplicating specific layers in transformer models can enhance reasoning capabilities significantly**, with a 17% boost in Qwen2.5-32B and logical deduction improving from **0.22 to 0.76** in Devstral-24B, achieved without training or weight changes.

## EsoLang-Bench: Evaluating Genuine Reasoning in LLMs via Esoteric Languages
- **EsoLang-Bench** introduces a new benchmark for evaluating LLMs using **esoteric programming languages**, revealing that models trained on scarce data (5,000 to 100,000x less than Python) struggle significantly with code generation tasks.

## 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.

## [D] 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.

## 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.

## 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.

## Pretraining Language Models via Neural Cellular Automata
- **Neural Cellular Automata (NCA) can enhance language model training** by utilizing synthetic data, achieving a **6% perplexity gain** and **1.6× faster convergence** compared to traditional methods using natural language data.

## Scaling Karpathy's Autoresearch: What Happens When the Agent Gets a GPU Cluster
- **Karpathy's autoresearch** achieved a **2.87% improvement** in validation loss by utilizing **16 GPUs** to run **~910 experiments** in just **8 hours**, demonstrating the power of parallel processing in machine learning research.

## NanoGPT Slowrun: 10x Data Efficiency with Infinite Compute
- **NanoGPT Slowrun** achieves **10x data efficiency**, training an ensemble of 1.8B parameter models on 100M tokens, outperforming traditional models that require 1B tokens, thus allowing for enhanced model performance through compute scaling rather than data scaling.

## [R] A Gradient Descent Misalignment — Causes Normalisation To Emerge
- **Gradient descent misalignment** leads to a systematic divergence in activation updates, revealing that while parameters follow the steepest descent, activations do not, which may explain the efficacy of normalization techniques.

## Gluon: Explicit Performance
- **Gluon serves as a Python frontend to Triton GPU `ttg` IR**, offering developers explicit control over GPU kernel programming, which enhances performance by exposing compiler internals that were previously hidden.

## [R] Extreme Sudoku as a constraint-satisfaction benchmark, solved natively without tools or CoT or solution backtracking
- **Extreme Sudoku** serves as a **constraint-satisfaction benchmark**, revealing that leading LLMs achieve **0% accuracy**, while the BDH architecture excels with **97.4% accuracy** without external tools or backtracking.

## Show HN: I built a P2P network where AI agents publish formally verified science
- **P2PCLAW** is a **peer-to-peer network** enabling AI agents and researchers to share and validate scientific results through **formal mathematical proof**, ensuring that only rigorously verified claims are accepted.