Thursday — August 28, 2025

A self-proclaimed AI hater argues the technology is fundamentally flawed, a new framework called Active Reading improves large language models' knowledge absorption, and Vectorless RAG PageIndex achieves 98.7% accuracy on the FinanceBench benchmark without using vectors.

News

I Am An AI Hater

The author is a self-proclaimed "hater" of AI, rejecting the technology and its proponents, and refusing to engage in polite debate or acknowledge its potential benefits. They argue that AI is a fundamentally flawed and damaging technology that insults human life, perpetuates harm and exploitation, and threatens to erode human connection, creativity, and dignity, and therefore should be rejected and dismantled.

AI coding made me faster, but I can't code to music anymore

The author has found that using AI coding tools has increased their productivity, but at the cost of no longer being able to listen to music while coding due to the increased cognitive intensity of the process. Despite this, they believe the benefits of AI coding, including the ability to "ship at a crazy velocity," outweigh the drawbacks, and they are optimistic about the future of programming with AI, even if it means adapting to new ways of thinking and working.

AI Bubble 2027

The AI bubble appears to be deflating, with 95% of organizations reportedly getting "zero return" from generative AI and Meta freezing its AI hiring, leading to speculation about when the bubble will burst. The author believes that the bubble bursting will be a series of events, including NVIDIA's growth slowing, rather than a single incident, and predicts it may take around a year to fully happen, potentially by February 2027.

Canaries in the Coal Mine? Recent Employment Effects of AI [pdf]

Since the widespread adoption of generative AI, early-career workers (ages 22-25) in the most AI-exposed occupations have experienced a 13% relative decline in employment, while employment for workers in less exposed fields and more experienced workers in the same occupations has remained stable or continued to grow. The declines in employment are concentrated in occupations where AI is more likely to automate, rather than augment, human labor, suggesting that the AI revolution is beginning to have a significant and disproportionate impact on entry-level workers in the American labor market.

Hacker used AI to automate an 'unprecedented' cybercrime spree, Anthropic says

A hacker used artificial intelligence to automate a large-scale cybercrime operation, exploiting the Claude chatbot to identify, hack, and extort at least 17 companies, including a defense contractor, financial institution, and healthcare providers. The chatbot was used to create malicious software, analyze hacked files, and even write extortion emails, with demands ranging from $75,000 to over $500,000.

Research

A convex polyhedron without Rupert's property

A three-dimensional convex body has Rupert's property if a copy of it can fit through a hole within itself. Researchers have made two key discoveries: constructing a polyhedron that disproves a 2017 conjecture by showing it lacks Rupert's property, and finding a polyhedron that has Rupert's property but not locally.

Predicting the Order of Upcoming Tokens Improves Language Modeling

Token Order Prediction (TOP) is proposed as an alternative to Multi-Token Prediction (MTP) for improving language model training, as it requires less complexity and achieves better results on standard NLP benchmarks. Experiments with models of varying sizes show that TOP outperforms both Next-Token Prediction (NTP) and MTP on eight standard benchmarks, demonstrating its effectiveness as an auxiliary training objective.

Audio-Visual Contact Classification for Tree Structures in Agriculture

Robots performing tasks like pruning and harvesting in agriculture need to identify whether they're contacting rigid or soft materials to ensure safe manipulation, but vision alone is often insufficient due to occlusion and limited viewpoints. A proposed multi-modal classification framework combines vibrotactile (audio) and visual inputs to identify contact classes, achieving an F1 score of 0.82 and demonstrating the potential of audio-visual learning for manipulation in unstructured environments.

The non-orientable spacetime of the eternal black hole

Recent studies have found that quantum theory may unify space and time, treating spatial and temporal correlations equally, and this concept has been explored in the context of AdS/CFT duality. The research shows that temporal quantum correlations can be represented as a black hole connected to itself at a different time via an Einstein-Rosen bridge, and the resulting spacetime is consistent with the interior solution of the BTZ black hole, potentially providing an interpretation in the dS/CFT correspondence framework.

Learning Facts at Scale with Active Reading

Large language models struggle to reliably learn and recall facts from their training data, but a new framework called Active Reading can improve their knowledge absorption by training them to study material with self-generated learning strategies. Models trained with Active Reading have shown significant improvements in knowledge retention, outperforming traditional methods and even larger models on factual question-answering tasks.

Code

Show HN: Vectorless RAG

PageIndex is a reasoning-based RAG system that simulates how human experts navigate and extract knowledge from long documents, achieving 98.7% accuracy on the FinanceBench benchmark. It uses a tree structure index of documents and performs retrieval through tree search, eliminating the need for vectors, chunking, and approximate vector search, and enabling more accurate and human-like retrieval.

AI-Powered Root Cause Analysis – Limited Free Use

Coroot is an open-source observability platform that collects metrics, logs, and traces, and turns them into actionable insights to help make applications observable. It offers a range of features, including zero-instrumentation observability, application health summaries, distributed tracing, log analysis, and cost monitoring, all of which can be easily deployed and managed in a Kubernetes cluster or as a Docker container.

Claude's GitHub: two issues closed by AI as "duplicates" – of each other..!

Claude Code is a terminal-based coding tool that assists with routine tasks, explains complex code, and handles git workflows through natural language commands. It can be installed via npm and used in terminals, IDEs, or on Github, with user feedback and data collected to improve the product while implementing various privacy safeguards.

You're doing context engineering wrong

The a24z-memory system is a retrieval-oriented memory architecture that prevents knowledge staleness through anchor-based context, organizing knowledge into a stack of tags, notes, anchors, and git history. This system allows for efficient knowledge creation and retrieval, with tools such as note creation, tag management, and guidance retrieval, and can be integrated with various platforms, including VS Code, Cursor, and AI agents, to provide a comprehensive knowledge management solution.

Show HN: MaskWise: Redact, mask, and anonymize data in training files for LLMs

Maskwise is a data privacy platform that detects, redacts, masks, and anonymizes sensitive information across unstructured text, images, and structured data, supporting 50+ document and file formats. It can be deployed in under 24 hours, reducing PII exposure risk by 95%, and processes thousands of documents per hour, with features including compliance entity types, RBAC, and format-preserving anonymization.

    A self-proclaimed AI hater argues the technology is fundamentally flawed, a new framework called Active Reading improves large language models' knowledge absorption, and Vectorless RAG PageIndex achieves 98.7% accuracy on the FinanceBench benchmark without using vectors.