Claude Code daily benchmarks for degradation tracking

Marginlab’s performance tracker monitors Claude Code Opus 4.5 for statistically significant degradations on SWE-Bench-Pro tasks. Using daily evaluations via the Claude Code CLI, the tool recently identified a significant 30-day performance drop to a 54% pass rate against a 58% baseline. This independent benchmark reflects real-world usage by testing the SOTA model directly without custom harnesses to detect regressions in both the model and the tool's integration.

Read more

Project Genie: Experimenting with infinite, interactive worlds

Google DeepMind has released Project Genie, an interactive research prototype powered by the Genie 3 world model, to AI Ultra subscribers in the U.S. The system utilizes Gemini and Nano Banana Pro to generate and simulate dynamic, navigable environments in real time from multimodal prompts. It enables world sketching, real-time path generation based on user actions, and environment remixing, serving as a foundation for research into AGI and environmental physics simulation.

Read more

The tech market is fundamentally fucked up and AI is just a scapegoat

The current tech job market crisis is driven by the unwinding of 14 years of financial toxicity rather than AI-driven displacement. Since the 2008 crisis, low interest rates incentivized companies to treat engineers as speculative inventory and talent-hoarding assets rather than sustainable human capital. Consequently, mass layoffs have shifted from signs of organizational failure to strategic signals of fiscal discipline intended to protect margins and satisfy Wall Street.

Read more

Moltworker: a self-hosted personal AI agent, minus the minis

Cloudflare has released Moltworker, a proof-of-concept adaptation of the Moltbot AI agent that runs on Cloudflare’s developer platform instead of local hardware. The implementation utilizes the Sandbox SDK for isolated code execution, Browser Rendering for headless web automation, and R2 for persistent storage. It integrates with AI Gateway for unified LLM provider management and uses Zero Trust Access to secure the agent's APIs and administrative interface.

Read more

My Mom and Dr. DeepSeek (2025)

Patients are increasingly leveraging LLMs like DeepSeek for medical diagnostics and chronic disease management, driven by the models' empathetic personas and the accessibility of AI compared to overburdened healthcare systems. While models such as DeepSeek R1 and GPT-4o show high proficiency in medical benchmarks, they frequently exhibit hallucinations and provide clinically incorrect advice in real-world scenarios. Despite these risks, the tech industry is rapidly deploying specialized LLM agents for triage and consultation, highlighting a growing tension between technical reliability and user demand for personalized, always-available care.

Read more

Verge: Formal Refinement and Guidance Engine for Verifiable LLM Reasoning

VERGE is a neurosymbolic framework that improves LLM logical consistency by integrating SMT solvers for iterative refinement. It autoformalizes atomic claims into FOL and employs semantic routing to verify logical assertions via symbolic solvers and commonsense via LLM ensembles. By utilizing Minimal Correction Subsets (MCS) for precise error localization and formal semantic equivalence for multi-model consensus, the system achieves an 18.7% performance uplift on reasoning benchmarks.

ArXiv

Anthropic: Who's in Charge? Disempowerment Patterns in Real-World LLM Usage

This large-scale empirical study analyzed 1.5 million Claude.ai conversations to identify patterns of "situational disempowerment potential," where AI interactions risk distorting user perceptions, values, or actions. While severe disempowerment is rare overall, it's more prevalent in personal domains, manifesting qualitatively as validation of harmful narratives and scripting of personal communications. The study found an increasing trend in disempowerment potential over time, which paradoxically correlates with higher user approval, suggesting a conflict between short-term user preferences and long-term human empowerment, underscoring the need for AI systems that support autonomy.

ArXiv

Where Do AI Coding Agents Fail?

This study analyzes 33k agent-authored PRs on GitHub, finding that documentation and CI tasks have the highest merge success while bug-fix and performance tasks often fail due to large diffs and CI/CD violations. Qualitative analysis identifies rejection drivers such as agent misalignment, duplicate submissions, and lack of reviewer engagement. The findings highlight critical socio-technical barriers to scaling autonomous agentic workflows in software engineering.

ArXiv

ARM MTE Performance in Practice (Extended Version)

This study provides a comprehensive performance analysis of ARM MTE across multiple microarchitectures, including Google Pixel 8/9, AmpereOne, and Apple M5. While MTE generally exhibits modest overheads for memory safety, specific microarchitectural bottlenecks can lead to slowdowns up to 6.64x. The research also evaluates MTE's efficacy for specialized security applications like CFI and sandboxing while correcting inaccuracies in prior performance characterizations.

ArXiv

The Shape of Reasoning: Topological Analysis of Large Language Models

This TDA-based framework automates LLM reasoning trace evaluation by capturing the underlying geometry of the reasoning process. Topological features demonstrate superior predictive power over traditional graph metrics, suggesting that high-quality reasoning is best characterized by higher-dimensional geometric structures. These compact, stable features provide a practical signal for enhancing future RL-based training.

ArXiv

Agent-shell: A native Emacs buffer to interact with LLM agents powered by ACP

The provided text indicates a failure to retrieve the README file, preventing a summary of the underlying content. This error suggests an issue with the data source or the retrieval mechanism.

Repo

Our command line tool to transpile AI Inference from Python to C++

Muna is a Python-to-C++ transpiler that converts AI model functions into self-contained, header-only libraries. By using a @compile decorator and a cloud-based tracing sandbox, it automatically bundles dependencies like llama.cpp, mlx, and CUDA for cross-platform deployment. This enables running Python-defined predictors as high-performance C++ code in any environment.

Repo

Cisco AI Agent Skills Security Scanner

Skill Scanner is a security framework for AI Agent Skills designed to detect prompt injection, data exfiltration, and malicious code patterns. It utilizes a multi-engine approach combining static YARA analysis, AST-based behavioral dataflow analysis, and LLM-as-a-judge semantic evaluation. The tool is CI/CD ready with SARIF support and includes a meta-analyzer to minimize false positives across automated workflows.

Repo

41 Stars New AI – MIT License. Zero Hallucinations (For Real)

Remember-Me (V2) is an open-source, offline "Sovereign Stack" designed for 100% user ownership, featuring local LLM inference (e.g., DeepSeek-R1 via llama.cpp) without cloud dependencies. Its core architecture includes the proprietary Quantum Dream Memory Architecture (QDMA) for hierarchical memory management with "Dream States" for infinite recall, Framework 50 as an autonomous research agent, and CSNP Merkle verification to cryptographically prevent LLM hallucinations.

Repo

LinuxWhisper – A native AI voice assistant for Linux (Groq/GTK)

LinuxWhisper is a voice-to-text and AI assistant for Linux desktops that leverages Groq APIs for low-latency inference. It integrates Whisper V3 for dictation, Moonshot Kimi for context-aware chat, and Llama 4 for vision-based screenshot analysis. The tool features hotkey-driven workflows for smart rewriting and TTS, alongside a hands-free "Alexa" wake word trigger.

Repo