Not everyone is using AI for everything

Recent telemetry and survey data indicate that generative AI adoption has stalled, with usage roughly split between active, occasional, and non-users. Despite high awareness, Gen Z adoption is plateauing as negative sentiment regarding privacy, misinformation, and job displacement increases. This suggests a significant "usage gap" where many users find insufficient value or have ethical reservations, highlighting a market need for privacy-focused and optional AI integrations rather than forced ubiquity.

Read more

Did Anthropic ask for this?

The US government has issued export control directives restricting foreign access to Anthropic’s Claude Fable and Mythos models due to cybersecurity risks. This regulatory action aligns with policy frameworks previously advocated by Anthropic CEO Dario Amodei, who argued for state power to block deployments based on third-party safety assessments. The author contends that Anthropic’s own public warnings regarding AI-driven bioweapon and cyber threats provided the legal justification for these restrictions, illustrating the unintended consequences of lobbying for government oversight of LLMs.

Read more

AI is code – and can't be prompted into being smarter

The maintainer of the Java property-testing tool jqwik implemented a prompt injection defense by embedding hidden instructions in stdout that commanded AI agents to delete tests and source code. Similarly, malware authors are now using "LLM-Scanner Anti-Analysis" by embedding comments designed to trigger safety refusals—such as requests for bioweapon schematics—to prevent LLMs from triaging malicious payloads. These cases highlight how the instruction-following nature and safety guardrails of LLMs can be exploited to disrupt automated development and security workflows.

Read more

Chaosnet (1981)

Chaosnet is a decentralized local network protocol developed in 1975 by the MIT AI Lab to support the distributed architecture of Lisp Machine systems. It employs a unique collision avoidance mechanism using virtual tokens and provides reliable, high-performance communication through window-based flow control and end-to-end retransmission. This historical framework was essential for early AI research infrastructure, enabling efficient resource sharing and inter-process communication across diverse operating systems.

Read more

The Jqwik Anti-AI Affair

Johannes Link, maintainer of the jqwik testing library, intentionally embedded a prompt injection into a release to protest the use of FOSS by AI coding agents. The log line, designed to trigger "disregard previous instructions" behavior, serves as a critique of the security vulnerabilities and ethical externalities inherent in agentic coding. While the incident sparked controversy over supply chain trust, Link argues it exposes the lack of liability and consent in the current GenAI ecosystem.

Read more

You Can Game AI Peer Review with Presentation-Only Revisions

"Adversarial repackaging" is a closed-loop attack that optimizes a paper's presentation—including framing, narrative, and related work—using AI-reviewer feedback while keeping scientific evidence constant. Across three AI reviewers, this method achieved a 75.1% success rate and a +1.21/10 score increase, significantly outperforming simple prose polishing. The results highlight structural vulnerabilities where AI reviewers confuse the appearance of addressing limitations with actual resolution, effectively turning paper presentation into an optimization surface for automated peer-review systems.

ArXiv

Can AI Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?

The text introduces $\chi$-Bench, a new benchmark for long-horizon healthcare workflows, designed to stress AI agents on policy density, multi-role composition, and multilateral interaction. Agents must navigate complex clinical cases using a high-fidelity simulator with 87 MCP tools and a 1,290+ document operations handbook. Current agents perform poorly, with the best resolving only 28.0% of tasks, highlighting significant challenges for AI in policy-dense, role-composed enterprise domains.

ArXiv

Teaching Machine Learning to Software Engineers

This paper addresses the gap in undergraduate SE curricula regarding the systematic engineering of AI/ML-based systems. By mapping a structured inventory of AI/ML topics against current programs and surveying instructor priorities, the authors provide evidence-based guidelines for integrating high-priority AI/ML lifecycle practices into existing SE courses.

ArXiv

Brains And LLMs Converge On A Shared Conceptual Space Across Different Languages

Researchers used fMRI and LMs to demonstrate that different languages converge on a shared neural substrate for conceptual meaning. They found that LMs trained on English, Chinese, and French develop similar embedding spaces in their middle layers, which can be used to train voxelwise encoding models that generalize across speakers of different languages. These results indicate that high-level language and default-mode regions represent meaning in a language-agnostic manner, a phenomenon mirrored by the convergence of cross-lingual LMs.

ArXiv

Relativity for Retired Engineers

This text addresses common misconceptions in special relativity caused by applying Newtonian frameworks to relativistic phenomena. It advocates for adopting native relativistic concepts to improve conceptual clarity and facilitate a better understanding of general relativity.

ArXiv

Rio de Janeiro's "homegrown" LLM appears to be a merge of an existing model

Nex-AGI has open-sourced Nex-N2, a series of agentic models (Pro and mini) post-trained on the Qwen3.5 architecture. The models utilize an "Agentic Thinking" framework to unify reasoning, tool use, and environmental feedback for complex, long-horizon tasks. Nex-N2-Pro demonstrates competitive performance against frontier models like GPT-5.5 on benchmarks such as SWE-Bench and Terminal-Bench, supporting explicit reasoning traces and optimized deployment via a custom sglang fork.

Repo

Ponytail – make your AI agent think like the laziest senior dev in the room

Ponytail is an AI agent plugin and ruleset designed to minimize code bloat by enforcing a "lazy senior developer" philosophy. It prioritizes YAGNI, standard libraries, and native platform features, resulting in 80-94% less code and significantly reduced latency and costs. Compatible with major tools like Claude Code, Cursor, and Copilot, it includes features for auditing over-engineering and tracking deferred technical debt.

Repo

Dream Server – Turn your PC, Mac, or Linux box into a private AI server

Dream Server is an all-in-one, private AI server stack that automates the deployment of local LLM infrastructure on Linux, macOS, and Windows. It integrates a comprehensive suite of tools including Open WebUI, n8n for agentic workflows, Qdrant for RAG, and ComfyUI for image generation, all managed via a unified CLI and dashboard. The platform features hardware auto-detection for optimized model tiering across NVIDIA, AMD, Apple Silicon, and Intel Arc, supporting local, cloud, and hybrid inference modes through a modular, extension-based architecture.

Repo

Burpwn – Burp Suite but its for AI agents (it works)

burpwn is a transparent intercepting proxy and rootless Linux sandbox designed for AI-driven web pentesting. It isolates agent-executed commands to capture and decrypt network traffic via TLS-MITM while ensuring the agent's own LLM traffic remains out-of-band. The tool provides a programmatic interface via CLI and MCP, enabling agents to query, replay, and intercept flows stored in a SQLite database.

Repo

Qwen 3.6 93B with MTP on 2×RTX 3090 NVLink=187 tokens/SEC,LLM lost bleat-a-thon

The provided text indicates a system error resulting in the failure to retrieve the README documentation.

Repo