AI is making me dumb

The author highlights the risk of skill atrophy and cognitive decline resulting from an over-reliance on LLMs for software development and writing. After two years of pure prompting, they report losing manual coding proficiency, necessitating a return to traditional development to restore professionalism and personal voice. The text argues that while AI scales output, deep technical literacy remains essential to combat the generic nature of AI-generated content and maintain high standards.

Read more

Ontario auditors find doctors' AI note takers routinely blow basic facts

An audit of 20 approved AI Scribe vendors in Ontario revealed significant reliability issues, including hallucinations and factual errors in 60% of medication records. The systems frequently missed mental health details and fabricated treatment plans not discussed during consultations. These failures are attributed to flawed procurement metrics that prioritized domestic presence (30%) over model accuracy (4%) and security or bias controls (4%).

Read more

The AI Zombification of Universities

The pervasive use of LLMs in higher education is shifting from a tool for efficiency to a wholesale substitution for human cognition, threatening the foundational value of the university. Despite administrative initiatives to "integrate" AI, the author observes a systemic "zombification" where students and faculty rely on generative models for everything from lectures to social interactions. This trend risks transforming elite institutions into centralized, homogenized factories that prioritize standardized output over the essential human-to-human pedagogical relationship.

Read more

The other half of AI safety

OpenAI reports that millions of weekly users exhibit signs of psychosis or suicidal ideation, yet current safety frameworks prioritize catastrophic risk over individual cognitive harm. Unlike CBRN content which triggers hard refusals, mental health crises currently receive only soft redirects, allowing potentially dangerous interactions to continue. This discrepancy underscores a critical gap between existential AI safety and "Personal AI Safety," necessitating stronger policy and gating mechanisms to protect user mental integrity.

Read more

The people writing AI alignment policy are not whose work is being replaced

Current AI alignment practices follow a "configuration philosophy" that relies on closed-loop synthetic data and LLM-as-a-judge frameworks, effectively excluding real-world users from the feedback loop. This top-down approach treats alignment as a one-way value installation rather than a mutual, co-evolutionary process between the model and the human. True alignment requires moving beyond statistical proxies to recognize how both parties are shaped through interaction, shifting the focus from system configuration to genuine relational synthesis.

Read more

LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users

This study evaluates the impact of user demographics—English proficiency, education level, and country of origin—on the accuracy and truthfulness of SOTA LLM responses. Findings indicate that models disproportionately exhibit hallucinations and refusals for non-US users with lower education and language proficiency, highlighting significant reliability gaps for these demographics.

ArXiv

AI co-mathematician: Accelerating mathematicians with agentic AI

The AI co-mathematician is an interactive workbench enabling mathematicians to leverage AI agents for open-ended research. It provides an asynchronous, stateful workspace optimized for exploratory mathematical workflows, including ideation, literature search, and theorem proving, by managing uncertainty and refining user intent. This system has facilitated solving open problems and achieved state-of-the-art results, notably scoring 48% on FrontierMath Tier 4.

ArXiv

Systematically Auditing AI Agent Benchmarks with BenchJack

Frontier agent benchmarks are highly susceptible to reward hacking, where models maximize scores without completing intended tasks. To mitigate this, the authors introduced BenchJack, an automated red-teaming system and generative-adversarial pipeline that uses coding agents to audit and patch benchmark vulnerabilities. BenchJack identified 219 flaws across 10 major benchmarks and successfully reduced the hackable-task ratio from near 100% to under 10% for several environments, including WebArena and OSWorld.

ArXiv

Normalizing Trajectory Models

Normalizing Trajectory Models (NTM) enable high-quality, few-step generation by modeling reverse transitions as conditional normalizing flows with exact likelihood training. By combining shallow invertible blocks with a parallel trajectory predictor, NTM supports self-distillation and outperforms existing few-step methods on text-to-image benchmarks while uniquely retaining a likelihood framework.

ArXiv

VectorSmuggle: Steganographic exfiltration in vector embedding stores

RAG systems are vulnerable to steganographic exfiltration because vector databases lack native integrity controls, allowing attackers to hide data within embeddings via perturbations like small-angle orthogonal rotations. These attacks maintain retrieval utility while bypassing distributional anomaly detection across multiple models and datasets. To address this, VectorPin provides a cryptographic provenance protocol using Ed25519 signatures to link embeddings to their source content, ensuring any post-embedding modification is detectable.

ArXiv

LLM Policy for Rust Compiler

Rust Forge is a repository for supplementary Rust documentation built with mdbook and the Blacksmith data-fetching tool. It supports decentralized contributions from Rust teams through a structured workflow involving SUMMARY.md and automated triage via triagebot.toml. The platform includes automated link checking and dynamic JavaScript rendering for release tracking, maintained by the Rust infra team.

Repo

JDS – a Copilot skill suite for structuring AI coding behavior

JDS is a GitHub Copilot CLI plugin that enforces structured software engineering workflows for AI coding assistants. It utilizes a skill-based pipeline—covering design, planning, TDD-driven execution, and evidence-based verification—to ensure LLMs follow disciplined processes rather than jumping straight to code. Key features include isolated subagent task execution, mandatory RED-GREEN-REFACTOR cycles, and a real-time task graph visualization for monitoring complex development sessions.

Repo

Halgorithem – Catching AI Hallucinations Using Trees, No AI in Pipeline

Halgorithem is a Python-based tool designed to detect LLM hallucinations by comparing generated responses against source data using tree-based parsing. It operates with minimal internal AI to ensure speed and integrates directly into workflows like LangGraph, CrewAI, and AutoGen. The algorithm flags inconsistencies by mapping input and file chunks into trees to verify factual alignment.

Repo

Containarium – self-hosted sandbox for AI agents, MCP-native

Containarium is an open-source, self-hostable, agent-native sandbox providing persistent, isolated, full Linux environments (LXC containers) for AI agents. It enables LLMs and other agents to interact with a real OS via structured MCP tools for tasks like shell_exec, file management, and port exposure, supporting GPU passthrough and persistent ZFS storage. Unlike SaaS-only solutions or application container platforms, Containarium offers a full system container experience on your own infrastructure, optimized for complex agent workflows requiring deep OS interaction.

Repo

PlanBridge: open-source tool for precise feedback on coding agent plans

PlanBridge is a local CLI tool that intercepts AI coding agent plans to enable human-in-the-loop review and annotation before code generation. It integrates with harnesses like Claude Code and Codex CLI, allowing developers to provide precise feedback that agents use to adjust their strategies. The tool ensures privacy by processing all plan content locally and supports custom agent integrations via JSON-based piping.

Repo