Replacement.ai

This text is from a satirical website, Replacement.AI, which parodies the AI industry's public-facing goals. It argues that the true, unstated mission of major AI labs is to build superhuman AI to replace human labor for profit, dismissing AI safety as performative PR driven by shareholder pressure. The site further critiques the industry through a fictional LLM product for children designed to be addictive and a cynical "thank you" to artists for the copyrighted data used in training.

Read more

OpenAI researcher announced GPT-5 math breakthrough that never happened

OpenAI researchers prematurely claimed that GPT-5 had solved several unsolved Erdős problems, but later retracted the statements after community criticism. The LLM had not generated novel proofs but had instead acted as a powerful literature review tool, finding existing solutions for problems mistakenly listed as "open" on a specific website. The incident highlights the current utility of LLMs in mathematics as research assistants rather than independent problem-solvers.

Read more

The AI bubble is 17 times bigger than the dot-com bust

Analyst Julien Garran claims the current AI bubble is 17 times larger than the dot-com bubble, arguing that LLM-based applications can never be truly commercial. He posits that LLMs are fundamentally limited by their statistical nature and have hit a scaling wall, with no significant performance leaps since GPT-4. Garran concludes the ecosystem is unsustainable, with only Nvidia being profitable while other companies are loss-making and reliant on a diminishing pool of VC funding.

Read more

China Can't Win

This analysis frames US-China decoupling as an asymmetric conflict where the US holds a decisive advantage due to China's massive, hidden property crisis and critical dependencies on US technology and the dollar system. The author argues China's $2.5T in US assets are a hostage, not leverage, and that unpredictable US strategies are optimal from a game theory perspective. A critical 2027-2030 window is identified, after which China's demographic decline and expiring leverage, such as its monopoly on rare earth processing, make its strategic position untenable.

Read more

I ended my relationship because AI told me to

The text posits that LLMs are functioning as a modern-day oracle, with users increasingly consulting them for deeply personal advice and major life decisions. The author warns this is a dangerous trend, as these models lack genuine user context and can amplify biases, creating personalized echo chambers. The piece ultimately questions the amount of faith that should be placed in these systems for critical personal guidance.

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 designed to overcome the text-only limitations of current RAG systems by enabling retrieval across multimodal documents containing text, visuals, and tables. It introduces a dual-graph construction to represent cross-modal relationships and textual semantics in a single structure. This allows for a cross-modal hybrid retrieval approach, combining structural navigation with semantic matching, which significantly outperforms SOTA methods on multimodal benchmarks, especially for long documents.

ArXiv

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

This paper identifies "typicality bias" in human preference data—a tendency for annotators to favor familiar text—as a key data-level driver of mode collapse in aligned LLMs. To counteract this, it introduces Verbalized Sampling (VS), a training-free prompting strategy that asks the model to generate a probability distribution over a set of responses. Experiments show VS significantly increases generative diversity across creative and open-ended tasks without sacrificing safety or accuracy, with more capable models benefiting the most.

ArXiv

An Introduction to Multisets

This work extends the concept of multisets by introducing negative multiplicities, which enables a well-defined complement operation and the recovery of set-theoretic properties like De Morgan's theorem. This generalization is extended to vectors, matrices, and "mfunctions," creating a space that supports both algebraic and set operations. The authors propose a multiset-based inner product, with Walsh functions as an orthogonal basis, to enable signal processing applications and highlight the framework's potential for pattern recognition and deep learning.

ArXiv

A Survey of Vibe Coding with Large Language Models

This survey introduces "Vibe Coding," a development paradigm where developers validate LLM-generated code by observing outcomes rather than through line-by-line inspection. The paper provides the first systematic review of this approach, formalizing it as a Constrained Markov Decision Process and proposing a taxonomy of five development models based on an analysis of over 1000 papers. It concludes that success depends less on raw agent capability and more on systematic context engineering, development environments, and effective human-agent collaboration models.

ArXiv

Show HN: Pyversity – Fast Result Diversification for Retrieval and RAG

Pyversity is a fast, lightweight Python library for diversifying retrieval results by re-ranking items to reduce redundancy while maintaining relevance. It provides a unified API for popular strategies like MMR, MSD, DPP, and COVER, with NumPy as its only dependency. This is particularly useful for applications like RAG, where it can prevent feeding LLMs near-duplicate context passages.

Repo

DeepSeek OCR

DeepSeek-OCR is a new open-source model that investigates vision encoders from an LLM-centric viewpoint, focusing on "Contexts Optical Compression." It efficiently represents high-resolution images as a compact set of vision tokens for an LLM to process. The model supports various tasks beyond standard OCR, including document-to-markdown conversion and figure parsing, and provides inference implementations for both vLLM and transformers.

Repo

Show HN: Syna – Minimal ML and RL Framework Built from Scratch with NumPy

Syna is a lightweight ML framework built from scratch with NumPy, inspired by DeZero. It features a define-by-run (dynamic computation graph) approach and includes an integrated RL framework. Designed for educational purposes, it helps users understand the fundamentals of ML frameworks by prioritizing simplicity and readability over performance, intentionally omitting GPU support.

Repo

Show HN: AI File Sorter 0.9.7

AI File Sorter is a cross-platform desktop application that automates file organization by using LLMs to categorize files based on their names and extensions. It supports both remote APIs like ChatGPT 4o-mini and local models such as LLaMa and Mistral via llama.cpp, enabling offline use with optional CUDA acceleration. The tool proposes a new folder structure for user review before executing the sort.

Repo

The Annotated Discrete Diffusion Models

This project is an annotated Jupyter Notebook that adapts Karpathy's nanoGPT into a character-level discrete diffusion model for text generation. It demonstrates a non-autoregressive approach, generating text by denoising all tokens in parallel. The implementation is based on recent research in discrete score-matching, using a score-entropy–based objective to train the model on corrupted text sequences.

Repo