# 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.

# MacBook M5 Pro and Qwen3.5 = Local AI Security System

- **Qwen3.5-9B achieves a remarkable 93.8% pass rate** on the HomeSec-Bench, only 4.1 points behind GPT-5.4, demonstrating the potential of local AI solutions on consumer hardware like the MacBook Pro M5.

# 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.

# 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.

# Flash-KMeans: Fast and Memory-Efficient Exact K-Means

- **Flash-KMeans** transforms the traditional **$k$-means** algorithm into an efficient online processing tool, addressing critical **IO bottlenecks** and **contention issues** in GPU implementations.

# 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.

# Attention Residuals

- **Attention Residuals (AttnRes)** enhances Transformer architectures by allowing layers to **selectively aggregate** earlier representations through learned attention, addressing the **dilution** of contributions in standard residual connections.

# Parallel Perl – autoparallelizing interpreter with JIT

- **Richard Jelinek** proposes a paradigm shift in programming by allowing **AI to write Perl**, leveraging decades of experience in AI and Perl development to enhance automation and efficiency in coding tasks.

# Launch HN: Canary (YC W26) – AI QA that understands your code

- **Canary** is an AI QA tool that intelligently analyzes code changes in pull requests, generating and executing tests to ensure user workflows function correctly, thus addressing the gap in pre-merge testing.

# [R] Doc-to-LoRA: Learning to Instantly Internalize Contexts from Sakana AI

- **Doc-to-LoRA (D2L)** introduces a **lightweight hypernetwork** that enables **instant context internalization** for Large Language Models (LLMs), allowing for efficient document understanding and multi-step reasoning without the need for extensive retraining.

# Medical AI gets 66% worse when you use automated labels for training, and the benchmark hides it! [R][P]

- **Medical AI performance declines by 66%** when trained with automated labels, revealing a significant risk in relying on such methods for segmentation tasks in breast cancer imaging.

# NumKong: 2'000 Mixed Precision Kernels for All

- **NumKong offers over 2,000 mixed-precision SIMD kernels** for various programming languages, optimizing performance across architectures like RISC-V, Intel AMX, and Arm SME, while maintaining a compact size of under 5 MB.

# [D] Tried MiniMax M2.7 impressive performance on real-world tasks

- **MiniMax M2.7 excels in handling complex tasks**, demonstrating impressive capabilities in coding workflows, bug tracing, and multi-step document edits, as experienced through ZenMux deployment.

# 🤗**Introducing SPEED-Bench: A Unified and Diverse Benchmark for Speculative Decoding**

- **SPEED-Bench** introduces a **comprehensive benchmark** for evaluating speculative decoding (SD) across diverse semantic domains and realistic serving conditions, addressing the limitations of existing benchmarks that often lack scale and diversity.

# NVIDIA GTC 2026: Live Updates on What’s Next in AI

- **NVIDIA** and **Thinking Machines Lab** have forged a **multiyear strategic partnership** to deploy **one gigawatt** of NVIDIA Vera Rubin systems, enhancing frontier model training capabilities.