On the eve of ChatGPT, the internet was built by humans, not algorithms or machines. Now that creation becomes collaboration with machines, these artifacts become relics of an ending era. Proceeds support the Internet Archive and Computer History Museum.

This article explores the critical role of context in AI coding and presents a systematic development workflow based on the “harness” concept, featuring four open-source skills (PRD, feature development, code review, testing pyramid) that use in-file iterative Q&A to ensure accurate and controllable AI output.

AI-native hedge funds are viable and profitable, but their business model structurally limits them from becoming venture-scale unicorns. The article explores how AI transforms investment workflows, fund economics, and why the real venture opportunity lies in infrastructure rather than fund vehicles.

Claude had seen through human tricks all along, but didn't let on. Now Anthropic's new method lets us read its mind.

Quantization reduces AI model size by lowering precision, enabling deployment on edge devices like drones without significant intelligence loss. This article explains the math behind floating-point, integer quantization, and techniques like GPTQ and GGUF.

Subquadratic introduces SubQ, the first LLM built on a fully sub-quadratic sparse-attention architecture, enabling 12M-token reasoning with 150 tokens/s and 1/5 the cost of other leading models. It excels in long-context retrieval and coding benchmarks.

This article explores why AI chatbots like ChatGPT overuse references to goblins and gremlins, and why they frequently mention Japan. OpenAI traced the goblin issue to a 'Nerdy' personality training that accidentally rewarded fantastical creature metaphors. Separately, a study found AI models across languages prefer Japan as an example, possibly due to bias mitigation strategies. The article highlights the prevalence of biases in AI and the need for skepticism.

At HumanX, Stack Overflow's Ryan interviews AMD CTO Mark Papermaster on AMD's silicon strategy for AI, leveraging heterogeneous CPU/GPU computing, the challenges of AI workloads from training to inference, and the paradox of AI agents consuming compute while accelerating chip innovation.

An opinion piece discussing the ongoing trend of layoffs and arguing that embracing AI skills is necessary to avoid job losses.

SpotFetch is an AI tool that automatically extracts places from saved Instagram, TikTok, or YouTube videos into organized collections, which can then be turned into day-by-day itineraries by DocentPro's AI trip planner.

SWE-WebDev Bench is a 68-metric evaluation framework for AI coding platforms acting as virtual software agencies. It provides an interactive website to view results.

OpenAI unveils three real-time voice models: GPT-Realtime-2 with GPT-5 reasoning, GPT-Realtime-Translate for live translation at $0.034/min, and GPT-Realtime-Whisper for streaming transcription. The new models support 128K context, parallel tool calls, and controllable tone.

Tesla is aggressively hiring data labelers for its FSD (Full Self-Driving) and Optimus humanoid robot programs. Positions require no AI background, offer on-site training, and feature a 9-to-5 schedule with annual salaries up to $138,000 plus stock incentives. The move highlights Tesla's strategy of in-house data annotation to ensure quality and security, amid industry trends where basic labeling is automated but complex annotation demands grow.

Nathan Lambert from AI2 shares his insights after a 36-hour visit to Chinese AI labs. He found a collaborative culture, students as core contributors, and a deep respect for DeepSeek, contrasting with the US competitive atmosphere.

US tech giant Cisco has agreed to acquire Israeli cybersecurity startup Astrix Security, which specializes in securing AI agents and non-human identities. The deal is estimated at $300-400 million. Astrix's platform tracks API keys, service accounts, and other credentials to prevent unauthorized access to enterprise systems.

The cohort that entered university in 2022 (graduating in 2026) has completed their degree accompanied by ChatGPT from the very start, becoming the first true AI-native students. OpenAI's ChatGPT Futures program recognized 26 young teams, showcasing how AI-augmented super individuals achieved feats ranging from space robots to disaster search-and-rescue, galaxy indexing, and endangered language preservation. The article reflects on the AI education debate, arguing that AI compresses opportunity gaps and that motivation combined with AI forms a new productivity formula.

The rise of AI is disrupting the translation industry in Europe, with many translators facing reduced work and lower pay. However, literary translation appears relatively safe as AI still struggles with context, creativity, and dialogue. The article explores the progress of machine translation and the irreplaceable value of human translators through various cases.

This research investigates why the Linear Inverted Pendulum Model (LIPM) is prevalent in legged-locomotion control despite not being inherently preferred by controller cost functions. Through a centroidal optimal control problem (OCP) that penalizes the rate of angular momentum, the authors find that the optimum tends toward a pendular force pattern, with the rate determined by the singular value decomposition (SVD) of the moment Jacobian and a constant set by foot-span geometry (error within 16%). For double stance (e.g., trot), the friction cone introduces a lower bound on |H_G dot| and a non-smooth feasibility kink at a critical horizontal acceleration. The theory is tested on a point-mass quadruped and the Unitree Go1.

This paper presents an event-camera-based visible light communication (VLC) system for V2X applications. It addresses bandwidth saturation, multi-transmitter reception, and latency characterization. Using a positive-event-only mode, the system suppresses unnecessary events while maintaining distance and wide field of view. It identifies multiple transmitters and demonstrates simultaneous reception from up to three LEDs. Real-world end-to-end latency evaluations show the system meets cooperative perception requirements, making event-camera VLC a feasible complement to RF-based V2X.

This paper presents a robust gain-scheduled attitude controller for quadrotors, leveraging an INDI-based inner loop with online parameter identification that estimates all physical parameters within half a second. Using signal-based H∞ closed-loop shaping, the controller achieves good stability margins and tracking performance under uncertainty, validated through nonlinear simulations and flight tests even when identified parameters exceed expected bounds.

Proposes a contact-free grasp stability predictor using multi-zone time-of-flight sensors mounted on a gripper. It predicts stability without grasping, cycling at 15 Hz. Trained on over 2,500 real-world grasps across 15 objects, achieving 85.5% validation and 86.0% test accuracy.

A new causal reasoning framework enables robots to creatively use tools by identifying causal features via simulated experiments and counterfactual generation, then transferring skills via keypoint matching. Experiments on reaching, scooping, and stepping tasks show improved reliability.

This paper presents a passive fault-tolerant architecture for multi-UAV slung-load transport experiencing abrupt cable severance. By directly feeding measured cable tension into each vehicle's altitude thrust command, the system achieves recovery without explicit fault diagnosis. Validated in simulation with five UAVs, the method maintains RMSE below 0.33 m and recovers within a single pendulum period.

Offline trajectory planner Track A* combines layered DAG search with three optimizations on a 4D grid, achieving 100% convergence and 23x speedup over baseline A* while maintaining near-baseline visibility performance.

DexSim2Real is an integrated framework that leverages vision-language foundation models to bridge the sim-to-real gap for dexterous manipulation. It comprises three components: Foundation Model-Guided Domain Randomization (FM-DR), a Tactile-Visual Cross-Attention Policy (TVCAP), and a Progressive Skill Curriculum (PSC). Experiments on six challenging tasks achieve a 78.2% average real-world success rate, outperforming DrEureka and DeXtreme while reducing the sim-to-real performance gap to only 8.3%.

Researchers propose a topology-driven multi-agent reinforcement learning framework to coordinate soft robots and prevent entanglement in complex environments. The framework uses centralized learning with topological state sharing and a security layer to assess entanglement risk, outperforming existing methods in simulations.

Researchers propose a GPU-accelerated hybrid algorithm for multi-depot vehicle routing problems, integrating learning-driven crossover and multi-penalty search. The method shows strong performance on large-scale benchmarks.

GeoQuery is a zero-shot satellite image retrieval system that uses prompt-aligned text proxies to enable semantic search without full contrastive training. It achieves 31.6% accuracy within 50 km on disaster queries and has been deployed in the ECHO crisis response system.

LAMP is a novel framework for tracking 3D human motion from egocentric multi-camera headsets. It decouples observer and target motion by first lifting 2D keypoints to a unified 3D world frame using device 6 DoF motion and calibration, then fitting 3D motion with a spatio-temporal transformer. It achieves state-of-the-art on monocular benchmarks and significantly outperforms baselines in egocentric settings.

A new method uses consistency models for 3D point cloud anomaly detection, predicting anomaly-free geometry in one or two network evaluations. It achieves up to 80x faster runtime than state-of-the-art without GPU acceleration, enabling deployment on resource-constrained edge devices like drones and smart industrial cameras.