> AI_ENGINEER.NEWS

FUTO Swipe collected over 1 million quality English swipe data, used to train and evaluate swipe typing models, now available on HuggingFace.

Baidu's Unlimited OCR enables one-shot, long-horizon parsing of images and PDFs using transformers, streamlining multi-page document processing.

Mistral OCR 4 offers multilingual, structured document extraction with bounding boxes, block types, and confidence scores, outperforming competitors.

VibeThinker-3B, a small model, surpasses larger models in verifiable reasoning, demonstrating advanced capabilities through optimized training.

Raymond, creator of the "Old New Thing" site and book, shares 30+ years of Windows history and anecdotes.

AI platforms heavily subsidize costs, risking massive debt, with prices rising to cover expenses, pushing AI companies toward displacing human jobs.

Anthropic launches Claude Tag on Slack, enabling collaborative, proactive task delegation and learning for teams, starting with beta access.

Multiple AI models experienced elevated errors; issue resolved and normal operations resumed after monitoring.

Elden Ring's NPC AI uses simple goal stacks and animation-driven actions, balancing efficiency and designer control over complex behaviors.

Lift4D improves 4D object reconstruction in challenging wild videos, handling occlusions and deformations with test-time optimization.

YOLO26 is a fast, edge-optimized multi-task vision model supporting detection, segmentation, pose, and classification, with improved accuracy and speed.

AI playing Civ reveals perception gaps, decision gaps, and safety risks in long-term strategic planning and goal tracking.

Neural Particle Automata expand NCA to dynamic particles using differentiable SPH, enabling scalable, self-organizing systems with new behaviors.

Ultralytics YOLO26 offers real-time, efficient detection, segmentation, and pose estimation with end-to-end architecture and improved accuracy.

AWS Lambda MicroVMs offer isolated, fast, resumable compute environments for multi-tenant apps, enhancing security and efficiency.

Modelplane is an open source control plane that manages AI inference deployments across cloud, on-premise, and multi-node setups.

LLMs fail to generalize reversed "A is B" patterns; workarounds are effective, but the reversal curse persists across models.

GLM-5.2 marks a major step for open AI agents, outperforming peers, intensifying competition, pricing, and regulatory challenges.

A detailed interactive knowledge graph maps AI and energy constraints, highlighting stress flows, chokepoints, and economic mechanisms.

AI coding agents now outperform humans in code review, making traditional review processes obsolete and scalable.

HALO uses RLMs and production traces to optimize AI agent harnesses, identify issues, and enhance performance through iterative self-improvement.

Modal Auto Endpoints enable users to own, control, and optimize LLM inference easily, with benchmarking, autoscaling, and transparency.

AI's business model is unsustainable with rising costs, collapsing tokenomics, and industry hype leading to price cuts and industry collapse risks.

AI's race for power mimics colonialism; large firms aim to make humans redundant, with rising global resistance pushing for regulation and accountability.

AI sparks existential fears; ChatGPT's rapid progress prompts reflection on meaning, productivity, and human value amidst technological change.

A GitHub tool rejects commits with AI co-author trailers, emphasizing accountability remains with human contributors.

An AI-driven legal firm helped a HR consultant win an English court case, marking the first trial won using AI legal assistance.

Agents reintroduce complex, stateful engineering challenges, requiring enhanced security, observability, and resilience for production deployment.

Built a cryptographically verified, fast AI memory engine in 10 days, enabling secure, predictable, and high-performance persistent memory.

LiteLLM switches to Rust, reducing latency to under 1ms and memory to 32MB, boosting throughput by 15x while maintaining compatibility.

A visual mini-course breaks down transformer building blocks interactively, making their core concepts accessible for all.

Built the fastest GLM-5.2 API at 280+ TPS using NVIDIA Blackwell, optimized inference, disaggregation, KV-aware routing, and speculation techniques.

Chinese AI startup Zhipu AI hits a trillion yuan, signaling strong state-backed growth and geopolitical ambitions in East Asia's AI race.

Diffusion models can generate high-quality samples without noise conditioning by leveraging high-dimensional geometry and stable parameterizations.

AI manages complex software development end-to-end, maintaining context across teams and systems for scalable, aligned engineering.

Graphsignal offers production-scale inference profiling to optimize AI performance, monitor errors, and identify bottlenecks across models and hardware.

Cachet is a Rust-based semantic cache for LLM APIs, boosting efficiency by caching rephrased prompts locally with a live savings dashboard.

GLM-5.2 outperforms Claude Opus 4.5 in benchmarks, offers larger context, costs 72% less, and suits most workloads better.

Verity.md gates, reviews, and remembers AI code changes pre-commit, offering security, quality, cost control, and smarter sessions.

Agent security requires a composition graph, not just package scans; understanding context reveals true risks and exposures.

A tiny, optimized semantic search model reduces size from 11.4MB to 2.79MB while outperforming larger models through domain-specific tricks.

SAA filters room speech, ensuring only directed voice commands reach your AI, reducing noise and improving interaction quality.

Open Knowledge Format improves knowledge portability but doesn't solve retrieval selection issues in AI tools; state management remains key.

Self-attention eliminated sequential bottlenecks, enabling parallel processing in transformers and vastly improving AI language model efficiency.

Fika Jobs raises $4M to build video-first AI interview platform, transforming hiring with AI-powered profiles and matching.

Linux Foundation proposes Agent Name Service, extending DNS to provide trusted, verifiable identities for AI agents across the web.

Google’s AI era challenges its dominance, with rising competition, user shifts away from AI, and talent departures impacting its search and innovation.

AI-agent activity in game development surpasses Twitter, highlighting areas with less AI adoption and potential FOMO risks.

Caplets streamlines AI agents with focused capability cards, reducing schema bulk, tool overload, and setup, enhancing efficiency.

A private, open-source pager enables AI agents to request human input securely via encrypted, accountless, relay-free connectivity.

OpenMontage is an open-source, agent-driven system that automates entire video production pipelines using AI tools and web search grounding.