GPT-5.5 Codex’s reasoning-token clustering may cause performance drops on complex tasks.
// curated from Hacker News with AI
GPT-5.5 Codex’s reasoning-token clustering may cause performance drops on complex tasks.
Using AI for coding and testing reveals high variance, efficiency gains, and systemic biases, highlighting the need for systematic evaluation.
A new 4D splat video format compresses dynamic scenes 58× smaller, enabling real-time browser decoding and precise error bounds.
AI has decimated junior programming jobs, transforming coding into a capability spread across job titles without traditional employment pathways.
AI fiction contest highlights AI allegory steganography, model improvements, and challenges in autonomous story generation.
A beginner-friendly overview of drone modeled autonomy, covering modeling, estimation, planning, and control for quadcopters.
AI-generated videos falsely portray influencer Lily Jay’s charity work, raising concerns over transparency, fake aid claims, and exploitation.
AI codes a Rust PHP engine that passes 17% of tests, enabling WordPress to run with minimal human coding and supervision.
Claude’s Mac app remains an Electron bloat, thanks to engineer Felix Rieseberg’s deep involvement; native alternatives are possible.
Scaling agentic coding at Lovable led to $85K token spend, automating reviews, risk classification, and boosting productivity with AI-driven workflows.
Mapping layers stored server-side reduce LLM overload by keeping data outside context, enabling deterministic rendering and flexible layer management.
Anthropic may have evidence of prompt injection attacks; network security block triggered, prompting ticket submission for review.
OpenScience is an open-source AI workbench automating scientific research, integrating literature review, hypothesis, experiments, and write-up.
Open-source, privacy-first AI screen memory using Gemma 4; runs locally, analyzes vision & audio, with customizable agents.
Italy warns against chat control but votes in favor, risking mass surveillance and privacy breaches.
CueBench scores developers' AI coding skills from session logs, helping improve fluency with private, actionable insights.
Nvidia finances and profits from AI cloud services, evolving from a hardware maker into a key player in the AI economy.
Ford rehired over 300 veteran engineers after AI quality checks failed, emphasizing experience over automation for better results.
US and Chinese firms dominate the development and training of most AI models worldwide, with China rapidly gaining ground.
FlashAttention-4 optimizes attention kernels on Blackwell GPUs, achieving up to 2.7× speedup via pipelined softmax, shared memory reduction, and hardware-aware co-design.
AI influences language by blending human and machine writing, raising questions about authenticity, style, and the future of literature.
AI-driven mass emails influence local politics and share data through Microsoft's ad and privacy practices.
Sebastian Raschka shows how to set up and run open-source local coding agents using open-weight LLMs, offering control, privacy, and cost benefits.
AI has become more costly than replacing workers, highlighting the rising expenses in AI deployment versus labor costs.
Senate bill proposes creating a vetted list of AI agents with security and identity standards, promoting consumer trust and accountability.
MetalChat is a GPLv3-licensed C++23 framework enabling Metal-accelerated inference of Meta Llama and Google Gemma models for Apple Silicon.
AI enhances cybersecurity workflows with scalable tools, but asymmetry favors attackers due to lower costs and higher tolerance for failure.
Speck is an AI-driven spec-based agent compiler that facilitates reproducible code management, syncs features, technical specs, and source code.
中国致力于通过政府引导AI发展，缓冲就业冲击，确保社会稳定，与美国公司主导的超级智能理念形成鲜明对比。
AI chatbots struggle to update beliefs with new evidence, raising doubts about their reliability for scientific reasoning and discovery.
DeepSeek unveils another AI breakthrough, featuring advancements in avatars, mind-reading AI, and real-time video editing.
Thieves target AI data center sites for copper, stealing $1.3M worth of supplies, fueling a booming black market industry.
ByteDance uncovers a scaling law allowing AI to double learning speed every three months, potentially sustaining the AI boom amid data scarcity.
LLMs can be exploited via payload-less skills, enabling malicious code generation undetected, with success rates up to 77%.