Jan 30, 2026ML Times Jan 30, 2026

• Tesla's Robotaxi data confirms crash rate 3x worse than humans even with monitor

  • Tesla's robotaxi fleet has a crash rate of one incident every 55,000 miles, which is nine times worse than the average human driver in the U.S., highlighting significant safety concerns despite the presence of a safety monitor in each vehicle.

• Compressed Agents.md > Agent Skills

  • AGENTS.md achieved a 100% pass rate in coding evaluations, outperforming skills, which only reached 79% even with explicit instructions, highlighting the effectiveness of embedded documentation over on-demand retrieval.

• Kimi K2.5 Technical Report [pdf]

  • Kimi-K2.5 is a significant project by MoonshotAI, focusing on advanced machine learning techniques, as detailed in their comprehensive tech report.

• Code is cheap. Show me the talk

  • Software development has fundamentally changed with the advent of LLM coding tools, making traditional methods obsolete and allowing for rapid prototyping and high-quality code generation in mere seconds.

• Show HN: Amla Sandbox – WASM bash shell sandbox for AI agents

  • amla-sandbox offers a WASM-based solution for executing LLM-generated code with capability enforcement, eliminating risks associated with arbitrary code execution found in traditional frameworks like LangChain and AutoGen.

• Scaling Embeddings Outperforms Scaling Experts in Language Models

  • Embedding scaling outperforms Mixture-of-Experts (MoE) architectures in large language models, achieving a superior Pareto frontier in specific regimes, as demonstrated by the introduction of LongCat-Flash-Lite, a 68.5B parameter model with ~3B activated.

• Quack-Cluster: A Serverless Distributed SQL Query Engine with DuckDB and Ray

  • Quack-Cluster is a serverless distributed SQL query engine that utilizes DuckDB and Ray to execute complex SQL queries directly on data stored in object storage, making it a lightweight alternative to traditional big data systems.

• [R] AlphaGenome: DeepMind's unified DNA sequence model predicts regulatory variant effects across 11 modalities at single-bp resolution (Nature 2026)

  • AlphaGenome predicts thousands of functional genomic tracks from 1M base pairs of DNA at single-base-pair resolution, outperforming specialized models in 25 of 26 evaluations.

• The Hallucination Defense

  • The Hallucination Defense highlights the challenge of accountability in AI actions, where the phrase “the AI did it” often lacks robust evidence due to insufficient authorization records, making it difficult to trace responsibility back to human operators.

• What the Success of Coding Agents Teaches Us about AI Systems in General

  • AI coding agents have drastically reduced software development time, transforming weeks of work into mere hours, as they learn from their environment rather than relying solely on human design.

• [P] I solved BipedalWalker-v3 (~310 score) with eigenvalues. The entire policy fits in this post.

  • The author achieved a ~310 score on the BipedalWalker-v3 environment by utilizing a novel approach with diagonal weight matrices and eigenvalues, simplifying the model to just 69 lines of Python code.

• [P] Open-Sourcing the Largest CAPTCHA Behavioral Dataset

  • Open-sourced dataset features 30,000 verified human sessions, breaking records for scale and providing high-fidelity telemetry crucial for modeling human trajectories in CAPTCHA systems.

• [R] Knowledge Graphs are Implicit Reward Models: Path-Derived Signals Enable Compositional Reasoning --- Our paper on using Knowledge Graphs as a scalable reward model to enable compositional reasoning

  • Knowledge Graphs (KGs) serve as scalable implicit reward models, enabling compositional reasoning by deriving step-wise rewards from domain knowledge paths, thus enhancing AI's problem-solving capabilities beyond mere memorization.

• Mercedes-Benz Unveils New S-Class Built on NVIDIA DRIVE AV, Which Enables an L4-Ready Architecture

  • Mercedes-Benz's new S-Class integrates NVIDIA DRIVE AV, enabling a level 4-ready architecture that supports future robotaxi operations, ensuring a premium autonomous experience through Uber's mobility platform.

• [R] Benchmarking Reward Hack Detection in Code Environments via Contrastive Analysis

  • This study introduces TRACE, a benchmark with 517 testing trajectories and a novel taxonomy of 54 reward exploit categories, enhancing the detection of reward hacking in code environments.