GPT-5.5

GPT-5.5 and GPT-5.5 Pro introduce enhanced agentic capabilities for coding, computer use, and scientific research, achieving SOTA results on benchmarks like Terminal-Bench 2.0 and SWE-Bench Pro. The models maintain GPT-5.4 per-token latency through infrastructure co-design with NVIDIA GB200 systems while offering improved token efficiency and a 1M context window for API deployments. Safety updates include stricter classifiers for cybersecurity and biology, alongside a "Trusted Access for Cyber" program for verified defensive work.

Read more

DeepSeek v4

DeepSeek provides API compatibility with OpenAI and Anthropic SDKs via dedicated base URLs, supporting models like deepseek-v4-flash and deepseek-v4-pro. Legacy model names deepseek-chat and deepseek-reasoner are scheduled for deprecation in July 2026 and currently map to the non-thinking and thinking modes of deepseek-v4-flash. The API supports standard chat completion parameters, including configurable reasoning effort and streaming options.

Read more

MeshCore development team splits over trademark dispute and AI-generated code

The MeshCore project has split following a dispute over the undisclosed use of Claude Code to "vibe code" major ecosystem components by a former team member. Citing community concerns regarding the reliability of AI-generated firmware and secret trademark filings, the core team has relocated to meshcore.io to focus on human-written software. This schism underscores growing tensions in open-source communities regarding the transparency and trust of LLM-integrated development workflows in low-level systems.

Read more

Our newsroom AI policy

Ars Technica’s AI policy mandates that all editorial content remains human-authored, prohibiting the use of generative AI for reporting, analysis, or primary creative assets. AI tools are permitted for research, data summarization, and workflow optimization, provided all outputs are human-verified and never used for direct source attribution. The framework emphasizes human accountability and requires clear labeling when synthetic media is used as an exemplar in AI-related coverage.

Read more

DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

DeepSeek-V4 introduces two MoE models, Pro (1.6T total/49B activated) and Flash (284B total/13B activated), featuring a 1M token context window. Key architectural innovations include a Hybrid Attention mechanism (CSA and HCA) that reduces KV cache requirements by 90% compared to DeepSeek-V3.2, Manifold-Constrained Hyper-Connections (mHC) for stable signal propagation, and the Muon optimizer. The models were pre-trained on 32T tokens and utilize a post-training pipeline involving domain-specific expert cultivation and on-policy distillation to achieve frontier-level performance in reasoning and coding.

Read more

Externalization in LLM Agents: Unified Review of Memory and Harness Engineering

LLM agent development is shifting from weight-based optimization to externalized runtime infrastructure that offloads cognitive burdens. By modularizing memory, skills, and interaction protocols into a unified harness, these systems transform complex tasks into more reliable operations. This framework suggests that future agent progress depends on the co-evolution of model capabilities and external cognitive infrastructure.

ArXiv

Sophia: A Scalable Second-Order Optimizer for Language Model Pre-Training

Sophia is a scalable second-order optimizer designed for LLM pre-training that utilizes a lightweight diagonal Hessian estimate and element-wise clipping. It achieves a 2x speed-up over Adam in total compute and wall-clock time across GPT models up to 1.5B parameters by significantly reducing the number of steps required for convergence. Theoretically, Sophia adapts to heterogeneous curvatures, offering a run-time bound independent of the loss condition number with negligible per-step overhead.

ArXiv

AI Assistance Reduces Persistence and Hurts Independent Performance

Current AI optimization for immediate task completion creates a short-sighted collaboration dynamic that reduces user persistence and impairs unassisted performance. Randomized controlled trials (N=1,222) reveal that even brief interactions lead to significant performance degradation when AI is removed, as users are conditioned to expect instant results. These findings suggest a need to shift AI development toward scaffolding long-term competence rather than just maximizing immediate output.

ArXiv

LLM users mistake AI output for their own real skill

This paper introduces the "LLM fallacy," a cognitive attribution error where users misattribute LLM-generated outputs to their own independent competence. By analyzing how low-friction, fluent interactions obscure the boundary between human and machine contributions, the authors establish a framework for understanding the divergence between perceived and actual capability. The work explores manifestations across computational and analytical domains, highlighting critical implications for education, hiring, and AI literacy.

ArXiv

MemReader: From Passive to Active Extraction for Long-Term Agent Memory

MemReader is a model family designed for active long-term memory extraction in agent systems, addressing the noise and inconsistency issues of passive, one-shot transcription. MemReader-0.6B is a distilled model for schema-consistent extraction, while MemReader-4B utilizes GRPO and a ReAct-style paradigm to perform reasoning-driven memory management, including selective writing, deferral, and context retrieval. Benchmarks on LOCOMO and HaluMem demonstrate SOTA performance in knowledge updating and hallucination reduction, with the models now integrated into MemOS.

ArXiv

Tolaria – Open-source macOS app to manage Markdown knowledge bases

Tolaria is an open-source, git-integrated markdown knowledge base for macOS designed to function as a local "second brain" or context repository for AI. It emphasizes data portability and offline-first principles, using plain markdown and YAML frontmatter to ensure compatibility with RAG workflows and AI agents. The platform is optimized for LLM integration, featuring an AGENTS file for navigation and support for tools like Claude Code and Codex CLI.

Repo

Run coding agents in microVM sandboxes instead of your host machine

SuperHQ is a Rust-based orchestration platform built with GPUI for running AI coding agents like Claude Code and Codex in isolated VM sandboxes. It utilizes a secure auth gateway reverse proxy to inject API credentials into outgoing requests, ensuring sensitive keys are never exposed to the sandbox environment. Key features include multi-agent support, port forwarding between host and VM, and a unified diff review panel for monitoring agent-driven file changes.

Repo

Cartoon Studio – an open-source desktop app for making 2D cartoon shows

Cartoon Studio is an open-source Electron desktop application that automates 2D cartoon production from scripts to MP4. It utilizes LLMs for script generation and vision-based mouth rigging of SVG assets created via Recraft V4. The pipeline integrates multiple TTS providers through a unified Speech SDK, employing Whisper-derived or native word timestamps for deterministic lip-syncing. Final rendering is achieved by composing HTML and GSAP timelines into video using HyperFrames, headless Chrome, and ffmpeg.

Repo

AgentBox – SDK to Run Claude Code, Codex, or OpenCode in Any Sandbox

AgentBox is a unified API for running coding agents like Claude Code and OpenCode within isolated sandboxes such as Docker, E2B, or Modal. It maintains full interactivity by running agents as server processes via WebSocket or HTTP rather than simple CLI wrappers. The SDK supports advanced LLM workflows including MCP server integration, multimodal inputs, sub-agent delegation, and custom sandbox image builds.

Repo

DecisionBox – Autonomous AI agent runs data discovery on your warehouse

DecisionBox is an open-source, autonomous AI agent platform designed for automated data discovery across major warehouses like Snowflake, BigQuery, and Redshift. It independently executes and validates 50–100+ SQL queries per run to generate ranked insights and recommendations, providing full transparency into its reasoning steps and SQL logs. The platform supports RAG-powered Q&A, semantic search via Qdrant, and integrates with various LLM providers including local models via Ollama.

Repo