Mar 7, 2026

• Hardening Firefox with Anthropic's Red Team
  AI models, like Claude Opus 4.6, have identified 22 vulnerabilities in Firefox, with 14 classified as high-severity, showcasing AI's potential to enhance software security rapidly. This collaboration with Mozilla highlights the effectiveness of AI in detecting vulnerabilities that could otherwise remain undetected for longer periods.

• Uploading Pirated Books via BitTorrent Qualifies as Fair Use, Meta Argues
  Meta argues that uploading pirated books via BitTorrent is fair use, claiming that this process is inherent to the technology and necessary for acquiring datasets crucial for AI training, thus framing it as a transformative act.

• Triplet Superconductor
  Scientists have potentially discovered a triplet superconductor in the alloy NbRe, which could enable ultra-fast quantum computers with minimal energy loss, marking a significant advancement in quantum technology.

• Sarvam 105B, the first competitive Indian open source LLM
  Sarvam 30B and 105B are open-source models designed for reasoning tasks, trained on 16 trillion and 12 trillion tokens respectively, showcasing strong performance in Indian languages and real-world applications, with Sarvam 105B excelling in complex reasoning and agentic workflows.

• LLM Doesn't Write Correct Code. It Writes Plausible Code
  LLMs produce plausible code that can be significantly slower than correct implementations, as demonstrated by a Rust rewrite of SQLite that is 20,171 times slower on basic operations despite passing tests and appearing functional.

• Filesystems Are Having a Moment
  Filesystems are gaining traction in AI as they provide a persistent context layer that enhances agent functionality, allowing for more efficient coding and project management without the complexity of traditional databases.

• A tool that removes censorship from open-weight LLMs
  OBLITERATUS is an advanced open-source toolkit designed to understand and eliminate refusal behaviors in large language models through a process called abliteration, which surgically removes internal representations without retraining.

• NanoJudge: Instead of prompting a big LLM once, it prompts a tiny LLM thousands of times.
  NanoJudge revolutionizes ranking by utilizing thousands of pairwise comparisons with tiny LLMs, yielding a mathematically rigorous leaderboard that overcomes traditional LLM limitations.

• Graph-Oriented Generation (GOG): Replacing Vector R.A.G. for Codebases with Deterministic AST Traversal (70% Average Token Reduction)
  Graph-Oriented Generation (GOG) replaces Vector RAG by utilizing a deterministic Symbolic Reasoning Model (SRM) to parse codebases as a Directed Acyclic Graph (DAG), achieving a 70% average token reduction in processing.

• Polar Factor Beyond Newton-Schulz – Fast Matrix Inverse Square Root
  Fast computation of the orthonormal polar factor for tall matrices is achieved through minimax polynomials and Jacobi preconditioning, enhancing the efficiency of the Muon optimizer in machine learning applications.

• Conversational LLM Evaluations in Minutes with NVIDIA NeMo Evaluator Agent Skills
  The nel-assistant skill streamlines LLM evaluations by enabling natural language configuration, eliminating the need for complex YAML files and manual debugging, thus enhancing developer efficiency.

• On-device speech toolkit for Apple Silicon — ASR, TTS, diarization, speech-to-speech, all in native Swift
  The open-source Swift package enables 11 speech models to run on Apple Silicon, utilizing MLX for GPU and CoreML for the Neural Engine, ensuring fully local inference without cloud reliance.