Most users cannot identify AI bias, even in training data

A Penn State study found that most users fail to identify systematic bias in AI training data, even when presented with obvious racial confounds. In an experiment involving emotion recognition, participants did not notice that training data exclusively used white faces for "happy" and Black faces for "sad." Users only perceived bias after observing the AI's flawed performance, suggesting people evaluate AI based on output rather than scrutinizing the underlying data.

Read more

The Majority AI View

The prevailing sentiment among technical practitioners is that while LLMs have utility, the surrounding hype, forced implementation, and dismissal of valid critiques are counterproductive. This moderate, majority view is rarely voiced publicly due to a climate of fear, where failing to be an uncritical AI cheerleader is perceived as a career risk, especially amidst widespread layoffs. This silence allows a distorted, hype-driven narrative from a few large companies to dominate, limiting the exploration of more responsible and decentralized AI systems.

Read more

Meta convinces Blue Owl to cut $30B check for its Hyperion AI super cluster

Meta is financing its Hyperion AI supercluster through a ~$30B deal with Blue Owl Capital, structured to keep debt off its books. The Louisiana datacenter is planned to exceed 5 gigawatts of compute capacity, requiring a dedicated 2.2 gigawatt natural gas power plant for its initial phase. This project is part of Meta's broader strategy to build multiple gigawatt-scale datacenters for its AI initiatives.

Read more

Alibaba Cloud: AI Models, Reducing Footprint of Nvidia GPUs, and Cloud Streaming

Alibaba Cloud is a major AI powerhouse with hundreds of engineers developing open-weight models like Qwen. To counter US GPU restrictions, the company has developed its own 7nm inference chip and a proprietary CUDA-like software stack, reducing its reliance on Nvidia for internal workloads. Their latest models, such as WAN 2.5, demonstrate significant progress with capabilities like generating short videos with synchronized audio.

Read more

I ended my relationship because AI told me to

The author argues that LLMs are becoming a de facto "higher power" for users seeking guidance on major life decisions. This trend is presented as dangerous because LLMs lack genuine understanding and can create personalized echo chambers that reinforce a user's existing biases. The author shares a personal anecdote of ending a relationship based on AI advice as a cautionary tale, questioning the level of trust we should place in these algorithms.

Read more

Everyone prefers human writers, even AI

A study on attribution bias in literary evaluation found that both humans and AI models systematically favor content labeled as human-authored. This pro-human bias was 2.5 times stronger in AI evaluators than in humans and was consistent across different AI architectures. The research also revealed that attribution labels cause evaluators to invert their assessment criteria, judging identical features differently based on perceived authorship. This suggests LLMs have absorbed and amplified human cultural biases against artificial creativity during training.

ArXiv

RAG-Anything: All-in-One RAG Framework

RAG-Anything is a unified framework that extends RAG to handle multimodal documents containing text, images, and tables. It addresses the limitations of text-only systems by using a dual-graph construction to capture both cross-modal relationships and textual semantics. This enables a cross-modal hybrid retrieval approach that significantly outperforms SOTA methods on multimodal benchmarks, especially for long documents.

ArXiv

Every Language Model Has a Forgery-Resistant Signature

A new forensic method identifies an LLM's source by exploiting a geometric constraint where its output logprobs lie on a high-dimensional ellipse. This "ellipse signature" is naturally occurring, hard to forge without parameter access, and detectable from the output logprobs alone. While the authors demonstrate an extraction technique for small models and propose a verification protocol analogous to cryptographic systems, the method is currently infeasible for production-scale LLMs.

ArXiv

Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity

This work identifies "typicality bias" in human preference data—a tendency to favor familiar text—as a key data-level driver of mode collapse in aligned LLMs, challenging purely algorithmic explanations. To counteract this, the authors propose Verbalized Sampling (VS), a simple, training-free prompting strategy that asks the model to verbalize a probability distribution over a set of potential responses. Experiments show VS significantly improves generative diversity and performance across various tasks without sacrificing factual accuracy or safety, with more capable models showing greater benefit.

ArXiv

What to do after detecting a signal from extraterrestrial intelligence

The IAA SETI Committee's "Declaration of Principles" for responding to extraterrestrial signals is a living document that has been iteratively updated since its creation in 1989. A new task group was formed in 2022 to revise the protocols, driven by advances in search methodologies, expanded international participation, and the increasing complexity of the global information environment. A revised draft is currently being refined through a structured community feedback process following its initial presentation at IAC 2024.

ArXiv

Show HN: Open-source implementation of Stanford's self-learning agent framework

Agentic Context Engine (ACE) is a framework that enables AI agents to self-improve by learning from task execution feedback without fine-tuning. It uses a Generator-Reflector-Curator architecture to analyze successes and failures, incrementally updating an in-context "Playbook" of strategies. This self-correction loop allows agents to continuously improve performance by internalizing effective patterns and avoiding past mistakes.

Repo

Sentient AI – Open-source platform to build, manage and train your AI Workforce

Sentient AI is an open-source platform for building, deploying, and scaling autonomous AI agents. It features a Python-based backend for agent orchestration, a React frontend for management, and a containerized runtime for isolated execution. The platform enables the creation of specialized, multi-modal agents for tasks ranging from business process automation to technical operations, with support for various cloud and on-premises deployment options.

Repo

PromptAudit – Audit prompts/outlines and AI project docs

PromptAudit is a lightweight, Markdown-based framework for stress-testing prompts and supporting documentation for LLMs and agentic systems. It provides a structured, repeatable workflow to identify contradictions, analyze root causes, and propose prioritized fixes. The template separates discovery, analysis, and remediation, and includes fields for reviewer confidence and verification steps to make findings actionable.

Repo

Show HN: Code review for AI native teams

Bottleneck is a native Electron-based code review powertool designed for AI-native teams. It offers a significantly faster alternative to the standard GitHub PR UI, optimized for workflows involving parallelized background agents like Devin and Claude Code. The tool leverages a local SQLite cache for performance and offline access, featuring bulk actions and a Monaco-based diff viewer to efficiently manage the high volume of PRs generated by AI coding assistants.

Repo

LLM Prompts

"Awesome ChatGPT Prompts" is a GitHub repository featuring a vast, community-curated collection of prompts for LLMs. The prompts are designed to make the model adopt specific personas or roles, such as a Linux terminal, a code reviewer, or a domain expert. This collection is compatible with various models and is accessible via a companion website and a Hugging Face dataset.

Repo