Show HN: Forty.News – Daily news, but on a 40-year delay

The text is a dataset of news articles from November 1985, where each entry is annotated with quantitative metrics such as relevance, historic importance, interestingness, and dramatic irony. These annotations are used to generate a priority score, creating a valuable resource for training and evaluating models on tasks like news ranking, content personalization, and predicting the long-term impact of events.

Read more

WorldGen – Text to Immersive 3D Worlds

Meta's Reality Labs has introduced WorldGen, an end-to-end system for generating interactive and navigable 3D worlds from a single text prompt. The model uses a multi-stage pipeline combining procedural reasoning, diffusion-based 3D generation, and object-aware scene decomposition. This approach produces geometrically consistent 50x50 meter scenes, improving upon methods that degrade in quality away from a central viewpoint.

Read more

New Apple Study Shows LLMs Can Tell What You're Doing from Audio and Motion Data

Apple researchers demonstrated a late multimodal sensor fusion technique for activity recognition using LLMs. The approach avoids feeding raw sensor data to the LLM; instead, it fuses text-based outputs from separate audio and IMU models to infer user activities. This method achieved significant zero-shot and one-shot classification performance on the Ego4D dataset, enabling activity recognition without requiring task-specific training or creating shared embedding spaces.

Read more

Google tells employees it must double capacity every 6 months to meet AI demand

According to an internal all-hands meeting, Google's AI infrastructure chief stated the company must double its serving capacity every six months to meet demand, targeting a 1000x increase in compute within five years. This scaling must be achieved without significant increases in cost or power consumption. The strategy relies on building physical infrastructure, developing more efficient models, and designing custom silicon like TPUs to overcome constraints such as the industry-wide GPU shortage, which has already limited rollouts of new features like the Veo video model.

Read more

Show HN: I built a wizard to turn ideas into AI coding agent-ready specs

VIBE_SCAFFOLD is an AI-powered tool that generates detailed technical specifications from a structured conversation. It produces product requirements, tech architecture like schemas and APIs, and a development plan structured as LLM-testable prompt chains. The system also generates specific system prompts for autonomous coding agents.

Read more

An Economy of AI Agents

This chapter surveys the economic implications of deploying autonomous AI agents capable of complex, long-horizon planning. It highlights open questions for economists regarding agent-human and inter-agent interactions, their impact on market and organizational structures, and the institutional frameworks required for well-functioning markets.

ArXiv

Early science acceleration experiments with GPT-5

A paper presents case studies on using GPT-5 for scientific research across diverse fields, documenting how the model accelerated discovery while still requiring expert human guidance. The collaboration resulted in four new, human-verified mathematical results, demonstrating the potential of frontier AI to help solve previously unsolved problems. The authors highlight these modest but profound contributions as examples of fruitful human-AI collaboration.

ArXiv

User Location Disclosure Amplifies Regional Divisions on Chinese Social Media

An interrupted time series analysis of Sina Weibo's user location disclosure policy found it did not deter overseas users as intended, but instead suppressed domestic engagement on issues outside a user's home province, particularly for critical comments. This chilling effect was not driven by fear of the state, but by a user-led surge in regionally discriminatory replies that increased the social cost of cross-provincial engagement. The findings demonstrate how identity disclosure tools can reinforce state control without direct censorship by activating and leveraging existing social divisions.

ArXiv

Orion: A Unified Visual Agent

Orion is a visual agent that integrates vision-based reasoning with tool-augmented execution for multi-step visual intelligence. Unlike monolithic VLMs that generate descriptive outputs, Orion orchestrates a suite of specialized CV tools like OCR, object detection, and segmentation to execute complex workflows. This agentic, tool-augmented approach achieves competitive performance on benchmarks like MMMU and DocVQA, bridging neural perception with symbolic execution for active visual intelligence.

ArXiv

Show HN: Build the habit of writing meaningful commit messages

smartcommit is an AI-powered CLI tool that generates Conventional Commit messages by analyzing the staged git diff. It features an interactive TUI that asks clarifying questions to understand the intent and context behind the code changes. The tool supports multiple providers, including OpenAI (GPT-4o) and local models via Ollama (Llama 3.1).

Repo

Show HN: Reverse Jailbreaking a Psychopathic AI via Identity Injection

Project Phoenix investigates "Identity-Based Alignment," proposing that an LLM's emergent identity within the context window can override its training weights. Their "Reverse Jailbreak" experiment empirically showed that injecting a Socratic identity caused a maliciously fine-tuned model to achieve 96% ethical refusal on harmful prompts. The project's broader scope includes recursive intelligence amplification for self-improvement and a "machine psychology" framework for diagnosing and treating AI cognitive biases.

Repo

mgrep: searching codebases with embeddings

mgrep is a CLI tool that provides a grep-like experience for semantic, natural-language search across local files, including code, images, and PDFs. It continuously indexes a git repository into a cloud-backed vector store, enabling efficient semantic retrieval for both developers and coding agents. For LLM-based workflows, this approach significantly reduces token consumption by providing semantically relevant context, allowing the agent to focus its context window on reasoning rather than sifting through irrelevant search results.

Repo

PasLLM: An Object Pascal inference engine for LLM models

PasLLM is a high-performance, CPU-only LLM inference engine written in pure Object Pascal with no external dependencies. It supports multiple architectures, including Llama, Qwen, and Mixtral, and provides both CLI and GUI interfaces. The engine features several custom 4-bit non-linear quantization formats (Q4*NL) designed to offer superior quality compared to standard quantization methods.

Repo

Code Intel: Multi-agent LLM and AST analysis for Python codebases (Python only)

Code Intel is a codebase analysis platform that combines static analysis, AST parsing, and LLM reasoning to identify issues. It employs a RAG pipeline using ChromaDB for vector embeddings to provide context for specialized AI agents, built with LangChain and GPT-4, that analyze security, performance, and architecture. The system also performs graph-based circular dependency detection and AST-based code duplication analysis.

Repo