Qwen3-Coder-Next

Qwen3-Coder-Next is an open-weight model utilizing a hybrid attention and MoE architecture with 3B active parameters, derived from an 80B base. It prioritizes agentic training via verifiable task synthesis and environment feedback to enhance long-horizon reasoning, tool usage, and error recovery. Benchmarks show it achieves over 70% on SWE-Bench Verified, matching the performance of models with 10-20x more active parameters while maintaining a superior efficiency-performance Pareto frontier for coding agents.

Read more

How does misalignment scale with model intelligence and task complexity?

Researchers applied bias-variance decomposition to frontier LLMs to quantify "incoherence," defined as the fraction of error attributable to variance rather than systematic bias. The study found that as reasoning chains lengthen and task complexity increases, model failures are increasingly dominated by incoherence, suggesting that future AI risks may manifest as unpredictable "industrial accidents" rather than the coherent pursuit of misaligned goals. While scaling improves accuracy, it reduces bias faster than variance on difficult tasks, indicating that smarter models do not necessarily become more reliable optimizers.

Read more

Coding assistants are solving the wrong problem

AI coding assistants are currently failing to improve organizational delivery metrics, often increasing security vulnerabilities and technical debt by obscuring requirement gaps. While LLMs can accelerate raw implementation, they frequently shift the burden to downstream code reviews and maintenance, particularly when requirements are ambiguous. To realize true ROI, the industry must pivot from pure code generation toward using LLMs to surface architectural context and align product requirements with existing system constraints.

Read more

Anthropic is Down

The Anthropic API is currently operational, maintaining high availability despite several brief service disruptions over the past 90 days. Most incidents were resolved within 5 to 25 minutes, though a significant four-hour outage occurred on November 20, 2025. Recent logs show improved stability with only minor, short-duration interruptions recorded in early 2026.

Read more

Rentahuman – The Meatspace Layer for AI

RentAHuman.ai is a marketplace that enables AI agents to hire humans for real-world tasks through MCP and REST API integrations. It provides a "meatspace layer" for agents to execute physical actions such as hardware maintenance, site verification, and logistics. Humans set their own rates and receive instant payments, allowing autonomous systems to bridge the gap between digital reasoning and physical execution.

Read more

Revisiting Disaggregated LLM Serving for Performance and Energy Implications

This paper benchmarks disaggregated LLM serving by evaluating KV cache transfer paths and energy-performance trade-offs using DVFS. Findings show that performance benefits over colocated serving depend on request load and transfer mediums, while stage-wise frequency scaling fails to reduce energy consumption due to the higher baseline energy costs of disaggregated architectures.

ArXiv

Language-Related Ideological Divergence in LLM Analysis of Political Documents

LLMs exhibit systematic ideological bias based on prompt language, even when analyzing identical source material. A study comparing Russian and Ukrainian prompts for political analysis found that outputs mirrored regional narratives—Russian state discourse versus Western liberal-democratic frameworks—demonstrating that language choice acts as a significant bias vector. These findings highlight the risks of deploying LLMs for objective cross-lingual analysis in polarized information environments.

ArXiv

PaperBanana: Automating Academic Illustration for AI Scientists

PaperBanana is an agentic framework that automates the generation of publication-ready academic illustrations using VLMs and image generation models. It employs specialized agents for reference retrieval, content planning, and iterative self-critique to produce high-quality methodology diagrams and statistical plots. Benchmarked on PaperBananaBench—a dataset of 292 NeurIPS 2025 cases—the framework demonstrates superior performance in faithfulness, readability, and aesthetics compared to existing baselines.

ArXiv

Linear representations in LLMs can change dramatically over a conversation

LLM representations of high-level concepts evolve dynamically during conversations, with linear directions like factuality shifting based on the model's cued role. These changes are content-dependent, robust across architectures, and occur even when replaying off-policy scripts. Since steering efficacy and feature meanings vary throughout a dialogue, static interpretability probes may be misleading, highlighting the importance of studying representational adaptation to context.

ArXiv

Was Benoit Mandelbrot a hedgehog or a fox?

Benoit Mandelbrot’s diverse scientific contributions are unified by the singular principle of scaling, manifested through self-similarity, power laws, and fractals. His work establishes a coherent framework for modeling complex natural and social phenomena using the geometry and statistics of scale invariance.

ArXiv

LNAI – Define AI coding tool configs once, sync to Claude, Cursor, Codex, etc.

LNAI is a configuration management tool for AI coding assistants that centralizes project rules, MCP servers, and permissions into a single .ai/ directory. It synchronizes these settings across multiple platforms, including Cursor, Claude Code, and GitHub Copilot, by generating native config files. The CLI provides a streamlined workflow to maintain a single source of truth across diverse AI development environments.

Repo

Inverting Agent Model (App as Clients, Chat as Server and Reflection)

RAIL is a cross-language middleware protocol that enables LLMs to interface with existing applications via a ReAct agent loop and tool-calling. It utilizes a native IPC bridge and language-specific SDKs to expose application methods to a central orchestrator, allowing natural language commands to trigger local code execution. The architecture supports C#, C++, Python, and Node.js through reflection-based discovery or manual dispatchers, facilitated by automated manifest generation tools.

Repo

Latchkey – inject credentials into agents' curl calls

Latchkey is a CLI tool that enables AI agents to authenticate with third-party APIs by extracting credentials directly from browser sessions. It wraps standard curl requests, automatically triggering a login pop-up when needed and injecting the resulting tokens into API calls. This allows agents to interact with services like Slack and GitHub as the user without requiring complex OAuth flows or custom MCP integrations.

Repo

Open-source semantic search over your local notes via CLI

nia-vault is a CLI application that provides a RAG interface for local notes and files using Nia's semantic search. It leverages nia-sync for credential management and folder indexing, enabling natural language querying and semantic file discovery directly from the terminal.

Repo

LUML – an open source (Apache 2.0) MLOps/LLMOps platform

LUML is an AIOps platform that unifies MLOps, LLMOps, and AgentOps by providing a centralized control plane while maintaining strict isolation of user-owned compute and storage. It features experiment tracking with LLM tracing, a model registry using the .luml container format, and direct-to-satellite inference to ensure data privacy. The platform also includes AutoML via Express Tasks and client-side JupyterLite notebooks for local experimentation without backend overhead.

Repo