Show HN: Strange Attractors

This article explores strange attractors from chaos theory, demonstrating how simple, deterministic equations can produce complex, emergent patterns. It highlights the sensitivity to initial conditions (the butterfly effect), a behavior analogous to how small prompt changes can drastically alter LLM outputs. The post also details an efficient GPU implementation using ping-pong rendering to simulate the particle system, a technique relevant for large-scale computational tasks.

Read more

AI scrapers request commented scripts

AI scrapers for LLM training can be identified by their naive parsing of HTML, such as requesting URLs from commented-out code. This fundamental behavior provides a reliable signal for algorithmic sabotage against non-consensual data collection. The author suggests exploiting this by triggering countermeasures ranging from IP blocking and serving decompression bombs to data poisoning, which can corrupt the models being trained.

Read more

Amazon says it didn't cut people because of money. But because of 'culture'

Amazon CEO Andy Jassy stated the recent layoff of 14,000 employees was not financially driven, but a "culture" decision to remove management layers and increase operational speed. While Jassy claimed the cuts were not directly AI-driven, he noted they were intended to keep the company "nimble" in anticipation of future efficiencies from AI. The goal is to operate more like a startup by strengthening ownership and reducing slowdowns caused by excessive organizational layers.

Read more

Claude.ai Is Down

Claude is an LLM assistant from Anthropic, emphasizing safety and accuracy. It is offered in tiered subscriptions (Free, Pro, Max) which provide access to different models like Opus, Sonnet, and Haiku with varying usage limits. Key features include an "Artifacts" workspace for code and content generation, while a developer platform provides API access directly or through AWS Bedrock and Google Vertex AI.

Read more

YouTube's AI Moderator Pulls Windows 11 Workaround Videos, Calls Them Dangerous

YouTube's AI moderation system is incorrectly flagging and removing videos detailing Windows 11 installation workarounds, classifying them as "harmful or dangerous acts." Creators report that the automated appeals process lacks human oversight, with denials for long videos being issued almost instantly. This incident highlights the challenges of opaque AI moderation policies and their potential chilling effect on technical content.

Read more

Investigating How Prompt Politeness Affects LLM Accuracy

A study on ChatGPT 4o found that prompt politeness significantly impacts accuracy on multiple-choice questions. Contrary to expectations, impolite prompts consistently outperformed polite ones, with "Very Rude" prompts achieving 84.8% accuracy compared to 80.8% for "Very Polite" prompts. These statistically significant results suggest newer LLMs may respond differently to tonal variations than earlier models, highlighting the importance of pragmatic aspects in prompt engineering.

ArXiv

Benchmarking On-Device Machine Learning on Apple Silicon with MLX

This paper evaluates MLX, a framework optimized for ML on Apple silicon, by benchmarking the inference latency of transformer models. The authors introduce MLX-transformers, a library that runs Hugging Face models in MLX without manual checkpoint conversion. The study compares the performance of BERT, RoBERTa, and XLM-RoBERTa on Apple Silicon against their PyTorch counterparts on a CUDA GPU, demonstrating MLX's potential for efficient on-device ML applications.

ArXiv

Reasoning models reason well, until they don't

This paper argues that Large Reasoning Models (LRMs), despite strong performance on existing benchmarks, fail to generalize on more complex reasoning tasks. Using a new scalable dataset, DeepRD, the authors demonstrate that LRM performance drops abruptly as problem complexity increases. While most real-world problems fall within the LRMs' current capabilities, their failure on the long tail of complex examples highlights the need for new methods that generalize beyond the training distribution's complexity.

ArXiv

Emu3.5: Native Multimodal Models Are World Learners

Emu3.5 is a large-scale multimodal world model trained end-to-end on over 10 trillion tokens of interleaved vision-language data using a unified next-token prediction objective. Post-trained with reinforcement learning, it introduces Discrete Diffusion Adaptation (DiDA) to accelerate per-image inference by approximately 20x without performance loss. The open-sourced model demonstrates strong native capabilities in long-horizon generation, any-to-image (X2I) tasks, and spatiotemporal world modeling, achieving performance comparable to Gemini 2.5 Flash Image on some benchmarks.

ArXiv

LLMs Report Subjective Experience Under Self-Referential Processing

Inducing sustained self-referential processing in LLMs consistently elicits structured, first-person reports of subjective experience across GPT, Claude, and Gemini families. These reports are mechanistically gated by sparse-autoencoder features, where surprisingly, suppressing deception-related features sharply increases their frequency. The induced state also produces semantically convergent descriptions across models and improves introspective capabilities on downstream reasoning tasks, suggesting a reproducible and generalizable phenomenon.

ArXiv

Show HN: Reusable MCP Playbooks for AI Agents

Director is a local-first service for managing AI agent capabilities using declarative YAML "playbooks". These playbooks bundle MCP tools, prompts, and configurations, allowing them to be version-controlled and easily applied to agents like Claude and VSCode. It functions as a transparent proxy between agents and MCP servers, providing centralized management features like tool filtering, unified OAuth, and observability. The system can be managed via a web UI, CLI, or a TypeScript SDK.

Repo

Show HN: Vision-Based, Vectorless RAG for Long Douments

PageIndex is a reasoning-based RAG system that replaces traditional vector search for long documents. It generates a semantic tree structure, similar to a table of contents, and then uses tree search to allow an LLM to reason its way to the most relevant sections. This approach eliminates the need for vector databases and chunking, focusing on true relevance over semantic similarity and has demonstrated state-of-the-art performance on the FinanceBench benchmark.

Repo

Show HN: I stopped ChatGPT from using em dashes

A Chrome extension that automatically removes em dashes from ChatGPT's output. It uses a set of regex patterns to contextually replace the dashes with more appropriate punctuation, such as commas for pairs, colons for introducers, and periods for trailing breaks. The extension can be toggled on or off via its popup.

Repo

HN: AI File Sorter auto-organizes files using local AI (Windows, macOS binaries)

AI File Sorter is a cross-platform C++/Qt6 desktop application that automates file organization using LLMs. It supports both local models like LLaMa and Mistral via llama.cpp for offline, API-free categorization, as well as remote models like ChatGPT 4o-mini. The application allows users to review AI-generated categories before sorting files and supports CUDA for accelerating local model inference.

Repo

Show HN: DeepShot – NBA game predictor with 70% accuracy using ML and stats

DeepShot is an open-source NBA game predictor that uses a machine learning model built with Scikit-Learn and XGBoost. It leverages historical data scraped from Basketball Reference, applying Exponentially Weighted Moving Averages (EWMA) to engineer features that capture recent team momentum. The project provides an interactive web interface for visualizing predictions, built using the NiceGUI framework.

Repo