AWS multiple services outage in us-east-1

A major AWS outage in US-EAST-1 was triggered by DNS resolution issues for DynamoDB. This led to a cascading failure, first impairing an internal EC2 subsystem for instance launches, which in turn broke Network Load Balancer health checks. The resulting widespread network connectivity issues impacted over 140 services, including Lambda, SQS, Bedrock, and SageMaker. Recovery involved throttling operations like EC2 instance launches and took over 12 hours to fully resolve.

Read more

Space Elevator

The text is a collection of short, disconnected factual snippets and labels, implicitly structured by increasing altitude. This non-narrative, multi-domain format presents a challenge for standard summarization, testing an LLM's ability to perform entity recognition and knowledge extraction. The input also simulates a multi-modal context where the model must rely solely on text descriptions of images.

Read more

Production RAG: what I learned from processing 5M+ documents

Based on experience building a production RAG system over millions of documents, the most impactful improvements came from advanced techniques beyond basic implementations. The highest ROI was achieved through LLM-based query expansion to generate multiple semantic and keyword queries, followed by a reranker to refine the retrieved chunks. Other key strategies included developing a custom, data-aware chunking logic, passing relevant metadata along with the chunk text to the LLM, and implementing a query router to handle non-RAG questions.

Read more

AI-generated 'poverty porn' fake images being used by aid agencies

NGOs are increasingly using generative AI to create images of extreme poverty for campaigns, citing lower costs and the ability to bypass consent issues. These synthetic images, often sourced from stock photo sites, are criticized for amplifying harmful racial and social stereotypes. This practice raises concerns about a feedback loop where biased AI-generated data is used to train future models, potentially amplifying prejudice at scale.

Read more

When a stadium adds AI to everything, it's worse experience for everyone

A blog post details a poor fan experience at a stadium after it implemented AI-powered, computer vision-based checkout systems for concessions. The author found these systems were significantly slower than human cashiers, causing long lines and still requiring staff intervention. The post speculates that the limitations of object recognition incentivized the stadium to drastically simplify its menus, reducing food variety and quality. The author concludes that this real-world AI application resulted in a degraded customer experience, contradicting vendor claims of increased efficiency.

Read more

Modeling Others' Minds as Code

The ROTE algorithm reframes human action understanding as a program synthesis problem, modeling social routines as behavioral programs rather than traditional policies. It leverages an LLM to synthesize a hypothesis space of these programs and uses probabilistic inference to reason over them from sparse observations. In experiments, ROTE significantly outperforms baselines like behavior cloning and other LLM-based methods in both in-sample accuracy and out-of-sample generalization.

ArXiv

Qwen Language Confusion Gate

This paper introduces the Language Confusion Gate (LCG), a lightweight, plug-in module that filters tokens during decoding to mitigate unintended language mixing in LLMs. Trained via norm-adjusted self-distillation, LCG predicts language families and applies masking to correct for sampling biases caused by differing output token embedding norms. The method significantly reduces language confusion across various models, often by an order of magnitude, without altering the base LLM or degrading task performance.

ArXiv

Reasoning with Sampling: Your Base Model Is Smarter Than You Think

This work introduces a training-free, MCMC-inspired iterative sampling algorithm that elicits advanced reasoning from base LLMs at inference time. By leveraging the model's own likelihoods, the method achieves performance comparable or superior to RL-posttrained models on benchmarks like MATH500 and GPQA. Unlike RL, this approach maintains sample diversity and requires no curated datasets or verifiers, suggesting broad applicability.

ArXiv

The Overfitted Brain: Dreams evolved to assist generalization

This paper proposes an "overfitted brain hypothesis," which uses a DNN analogy to explain the function of dreams. It posits that the brain, like a neural network, is prone to overfitting from daily learning. Dreams act as a biological regularization technique, similar to noise injection, by creating corrupted sensory inputs from stochastic neural activity. This process is theorized to combat overfitting and improve the brain's ability to generalize.

ArXiv

The Circular Electron Positron Collider (CEPC) Technical Design Report

The Circular Electron Positron Collider (CEPC) is a proposed 100-km particle collider designed as a Higgs factory to generate a massive dataset for high-precision physics, including millions of Higgs bosons and trillions of Z bosons. This text is the second volume of the CEPC's Technical Design Report (TDR), detailing the architecture, performance, and cost of the reference detector designed to capture this data. The report also covers alternative detector concepts and outlines a project timeline targeting data collection in the 2030s.

ArXiv

DeepSeek OCR

DeepSeek-OCR is a model designed for "Contexts Optical Compression," investigating vision encoders from an LLM-centric viewpoint. It processes images and documents into a compact set of vision tokens for tasks like markdown conversion, figure parsing, and general OCR. The model supports various native and dynamic resolutions and provides inference implementations for both vLLM and Transformers.

Repo

Show HN: Visual autocomplete for drawings (real-time Human-AI interaction)

AIDrawing is a C++ application that provides real-time, AI-powered drawing assistance without requiring text prompts. It uses an Ollama-based vision model to interpret the user's strokes and generate a semantic description of the canvas. This description, along with the current drawing, is fed into a StreamDiffusion pipeline for real-time image-to-image generation. The final output is composited back into the application via Spout, creating a closed-loop system for AI-driven auto-completion.

Repo

Show HN: Starbase – Browser-based MCP server testing with AI chat integration

Starbase is an interactive tool for developing, testing, and debugging MCP servers. It addresses the difficulty of validating MCP tool integrations by allowing developers to connect to any remote server with just a URL. This provides an immediate, batteries-included playground to test the server's functionality against various LLMs, including Claude and GPT-4, without complex setup.

Repo

Show HN: Hank – Simplest CLI tool to get errors in plain English

hank is a simple CLI utility that uses a local LLM to provide clear explanations for program errors. It functions as a command wrapper, intercepting and analyzing stderr when you prepend hank to your command (e.g., hank make all). The tool is designed for privacy-conscious developers as it relies exclusively on a locally served model via LMStudio, offering a lightweight alternative to full agentic terminal solutions.

Repo

Show HN: ContextKey – Use a hotkey to query LLM using any text or file

ContextKey is a native macOS menu bar app for querying LLMs on any selected text, image, or file via a hotkey. It supports local models through Ollama and any remote API via a flexible custom endpoint configuration. Users can define the full HTTP request template and specify the JSON response path for parsing, allowing integration with services like OpenAI, Anthropic, or custom backends. The open-source tool is privacy-focused, storing all conversation data and API keys locally.

Repo