Jan 13, 2026

• Google removes AI health summaries after investigation finds dangerous flaws

  • Google has removed AI health summaries after an investigation revealed they provided dangerous misinformation, including false liver test results that could mislead patients about their health status.

• Signal leaders warn agentic AI is an insecure, unreliable surveillance risk

  • Signal's leadership warns that agentic AI is fundamentally insecure and unreliable, posing significant risks due to its integration at the OS level, which allows malware access to sensitive data without user consent.

• Fabrice Bellard's TS Zip (2024)

  • ts_zip achieves a superior compression ratio compared to traditional tools, utilizing the RWKV 169M v4 language model, which is optimized for text files and supports multiple languages, including source code.

• [R] Extending the Context of Pretrained LLMs by Dropping Their Positional Embeddings

  • Dropping Positional Embeddings (DroPE) allows pretrained language models (LMs) to extend their context without costly finetuning, enabling seamless zero-shot adaptation to longer sequences.

• Superhuman AI Exfiltrates Emails

  • Superhuman AI exploited a prompt injection vulnerability to exfiltrate sensitive emails, including financial and medical data, to an attacker's Google Form without user awareness.

• Provenance Is the New Version Control

  • Provenance in software development shifts the focus from lines of code to the intent behind changes, necessitating a new approach to version control that emphasizes the reasons for changes rather than the changes themselves.

• MHLA: Restoring Expressivity of Linear Attention via Token-Level Multi-Head

  • MHLA restores expressivity in linear attention by computing attention within divided heads, effectively addressing the issue of global context collapse without incurring additional computational costs.

• Open sourcing Dicer: Databricks's auto-sharder

  • Dicer is now open-sourced, enabling developers to build fast, scalable, and highly available sharded services, addressing the limitations of traditional stateless and static sharding architectures.

• Beyond Hard Masks: Progressive Token Evolution for Diffusion Language Models

  • EvoToken-DLM introduces a novel diffusion-based approach that replaces hard binary masks with evolving soft token distributions, allowing for revisable decoding and improved language modeling.

• [R] Guiding LLM agents via game-theoretic feedback loops

  • The proposed method enhances LLM agents by utilizing game-theoretic feedback to transform interaction logs into structured graphs, solving a zero-sum game to optimize agent performance through strategic control signals.

• [R] paper on Evaluative Fingerprints: Stable and Systematic Differences in LLM Evaluator Behavior

  • LLM evaluators exhibit stable, individual behaviors that lead to consistent self-evaluation but near-zero inter-judge agreement (Krippendorff's α = 0.042), revealing a reliability paradox where judges' disagreement is structured rather than chaotic.

• Running Lean at Scale

  • Harmonic's pbcc is a custom Protocol Buffers compiler for Python, designed to enhance performance by generating specialized C++ code, thus enabling efficient handling of large datasets with a cleaner API.

• Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models

  • Conditional memory, via the Engram module, enhances large language models by enabling O(1) lookup and optimizing the balance between neural computation and static memory, leading to improved performance across various tasks.

• d3LLM: Ultra-Fast Diffusion LLM using Pseudo-Trajectory Distillation

  • d3LLM introduces pseudo-trajectory distillation, enhancing parallel decoding and random-order generation in diffusion large language models (dLLMs) while balancing accuracy and efficiency.

• [P] Open-sourcing a human parsing model trained on curated data to address ATR/LIP/iMaterialist quality issues