Every Agent Needs a 'Box': Box Founder
Every Agent Needs a 'Box': Box Founder
Box founder Aaron Levie notes strong industry discussion around code review collaboration, suggesting growing demand for efficient coding tools. He encourages readers to review related content about improving development workflows.
Anti-Patterns in Agent Engineering
Simon Willison identifies common anti-patterns in agent engineering to avoid, such as submitting unreviewed code to collaborators. He emphasizes developers should never submit pull requests without their own review, calling this a widespread and frustrating practice.
Federal Inference: Privacy-Preserving Model Collaboration
An arXiv paper introduces Federal Inference (FI) technology, allowing multiple independently trained models to collaborate during inference without sharing data or parameters. This research addresses privacy issues in distributed inference through encrypted protocols ensuring data security. Experiments show FI reduces data leakage risks by 90% while maintaining model performance, making it suitable for sensitive fields like healthcare and finance.
GPT-5.2 Pro Derives Graviton Amplitudes
OpenAI’s new preprint extends single-amplitude theory to gravitons, with GPT-5.2 Pro helping derive and validate non-zero graviton tree amplitudes in quantum gravity. This breakthrough could provide new tools for unified field theory.
Zipf-Preserving Alternative Model for Sequences
arXiv paper introduces a novel symbolic sequence alternative model that preserves Zipf’s law distribution and long-range correlations in language and DNA data. The model shows potential applications in natural language processing and bioinformatics fields.
RxnNano Trains Compact LLMs for Chemistry
arXiv paper ‘RxnNano’ proposes training compact LLMs through curriculum learning for chemical reaction and retrosynthesis prediction. The model outperforms existing parameter-heavy methods in drug discovery tasks while reducing parameters by 30%.
LangSmith Launches CLI and Skill Suite
LangChain released CLI tools and first skill suites, adding agent tracking, execution analysis, and performance evaluation to the LangSmith ecosystem. These features boosted Claude Code’s performance by 15% in testing sets, enhancing AI coding agents’ professional capabilities.
Anthropic $19B Revenue, Qwen Team Shakeup
Anthropic’s annual revenue reached $19 billion as Qwen’s core members departed, while Gemini and GPT recently updated with faster model versions. Market data shows intensified competition in large models as leading companies accelerate iterations.
Qwen Team Changes Spark Concerns
Simon Willison noted multiple Qwen team members left within 24 hours after releasing the Qwen 3.5 open-source series. He worries the 3.5 models might be the team’s final work, suggesting future uncertainty despite recent open-source contributions.
ERI Benchmarks Engineering Model Capabilities
arXiv’s Engineering Reasoning and Instruction benchmark released the first engineering instruction dataset covering 9 disciplines (e.g., civil engineering) for training and evaluating capable LLMs and agents. It includes 5,000 complex instructions for engineering domain evaluation.
Meta's NLLB-200 Shows Universal Concept Structures
Meta studied the NLLB-200 model (200 languages) to analyze if neural MT models learn language-agnostic concepts. The model exhibits cross-lingual concept mapping rather than surface-level clustering, achieving 0.78 accuracy in semantic similarity tasks—outperforming language-family baselines. This research informs multilingual model design for better cross-lingual reasoning.
LangChain Launches Skill Library Boosting AI Coding
LangChain released the first skill library with open-source tools (LangChain, LangGraph, Deep Agents) that enhanced Claude Code’s task completion rate from 29% to 95%. The open library enables developers to quickly build AI agents with professional coding capabilities, with tests showing integrated agents independently completed 85% of code generation tasks.
ATPO Algorithm Boosts Medical Dialogue Accuracy
ArXiv introduces the ATPO algorithm to optimize information retrieval in multi-turn medical dialogues. It addresses incomplete information in diagnosis by enhancing LLM interaction through tree strategy optimization. ATPO improves accuracy by 15% in medical QA tasks and can be integrated into clinical dialogue systems for developers.
MoE Models Need Router Calibration for Efficient Compression
Research shows MoE models face memory bottlenecks during deployment despite efficient scaling. The team introduces three compression paradigms—expert pruning, editing, and sharing—without retraining. Router calibration improves MoE inference speed by 40% while maintaining 92% performance, optimizing resource utilization for large-scale AI model deployment.
SuperLocalMemory: Secure Multi-Agent Memory System
arXiv paper introduces SuperLocalMemory, a memory system for multi-agent AI using architectural isolation and Bayesian trust scoring to defend against memory poisoning. It achieves 40% lower error rates in OWASP ASI06 attack scenarios through adaptive learning, suitable for sensitive data applications.
Memory Extraction in Diffusion Language Models
Diffusion language models (DLMs) differ from autoregressive models (ARMs) in memory handling. Research shows DLMs struggle with direct memory extraction but may expose memorized information during sampling. This finding has implications for model copyright and privacy, requiring developers to focus on data cleaning.