LIBERO-MAX evaluates 14 robot policies across 8,000 paired scenarios and 224,000 simulations, showing all policies lose 11–25.7 percentage points in success rate when environments change mid-execution, and even higher observation frequency fails to recover the gap.
ZeroClaw rewrites OpenClaw in Rust, delivering a 5MB single binary that starts in under 10ms and uses ~16MB idle memory, enabling AI agents on Raspberry Pi and low-end VPS with default supervised security, four OS-level sandboxes, and direct GPIO/I2C/SPI/USB hardware control via a Peripheral trait.
This article explains why traditional monitoring fails for AI agents and introduces AgentOps—covering observability with full traces, five key SLOs (task success rate, human escalation rate, cost per successful task), a multi-stage release pipeline with shadow and canary phases, FinOps with model routing, ROI calculation including risk costs, agent lifecycle management, and a continuous improvement flywheel turning production traces into eval datasets.
Anthropic's Claude Haiku 5.5 surpasses DeepSeek V4.1 Flash and GLM-5.3-Flash to lead small models, matching GPT-6 Luna on benchmarks while cutting prices 90% for sub-100k token requests, adding five-tier effort control, and improving computer-use accuracy to 72.4% on OSWorld 2.1, though tokenizer changes may increase token consumption and it cannot replace Sonnet 5.5 for complex coding.
This article explains the concept of an agent harness—a user-owned software layer that provides environment, system prompts, tools, an agentic loop, and a translation layer for AI models—using a climbing harness analogy, and argues that open-source, neutral harnesses like Pi empower user autonomy over AI.
DeepSeek Bot (ds bot) is a community-developed open-source plugin for DeepSeek Harness that creates persistent AI agents with independent memory, coordinated by a chief bot that manages group chats, task delegation, and model configuration, enabling multi-agent collaboration with local data storage and usage tracking.
DACon 2026 Beijing showcases 14 presentations from companies like Ant Group, Li Auto, and Shopee detailing how they successfully deployed ontology in production systems, covering architectural choices, semantic layer engineering, knowledge graph integration, and measurable business outcomes across finance, automotive, and e-commerce.
This article explains how to combine RAG with AI agents to create Agentic RAG, covering four collaboration patterns (RAG as Tool, Agent as Router, Multi-step RAG, Self-RAG), practical Java Spring Boot implementation with retrieval tools, self-evaluation loops, multi-step reasoning, architecture diagrams, and practical tips on iteration limits and cost monitoring.
LeCun's startup AMI releases H-JEPA, a hierarchical world model that stacks JEPA layers to plan at multiple time scales, achieving 73% success on Visual AntMaze versus 18% for single-layer models, with code and pre-trained weights fully open-sourced.
Zhejiang University and Alibaba propose Relay-OPD, an online distillation method that detects student reasoning errors via teacher-student token divergence and briefly hands control to the teacher for targeted correction, achieving +5.73% accuracy on math benchmarks while halving training trajectory length.
OpenAI adds Ultrafast tier to GPT-6.1 Sol API at 6x standard price for latency-sensitive workloads; benchmarks reveal model matches GPT-6 Astra on coding and agent tasks at far lower cost, with detailed guidance on caching, routing, and production safety for developers.
This article argues that AI agents often mistake model output or tool calls for task completion, but real business completion requires verifiable read-back evidence, idempotency handling for timeouts, explicit acceptance criteria, and structured completion records — not just a 'final_output' signal.
Meta FAIR veteran and Recursive Superintelligence co-founder Tian Yuandong recounts his journey from AlexNet debates to Go AI, Coconut, and grokking research, revealing how co-authoring a paper with GPT-5 boosted efficiency 6-10x and led him to predict AI will replace researchers within five years, while arguing humans must still initiate and verify ideas.
Google is internally testing a new Gemini 4 checkpoint codenamed Carbon, which employees say matches Anthropic's Opus 5.5 for coding, while the already-announced Argon model remains unavailable to most users despite strong benchmarks in long-context software engineering but weaknesses in terminal interaction.
Google DeepMind launches Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber with lower token usage, cheaper pricing, and specialized capabilities, while confirming the most aggressive pre-training ever for Gemini 4.
Tencent Multimedia Lab and Google collaborated at IEEE ICIP 2026, presenting nine research papers on AV2 video coding standard tools including intra prediction, block partitioning, inter coding, transform, warp prediction, in-loop filtering, entropy coding, screen content coding, and a rate-distortion optimization framework for 3D mesh geometry compression.
Google's TimesFM-3 introduces native multivariate time series forecasting with alternating causal temporal and full variate attention, non-autoregressive decoding via contiguous patch masking, and support for historical and future-known covariates, achieving top rankings on Gift-Eval, FEV-Bench, and TIME benchmarks without sacrificing univariate performance.
Li Auto engineers reveal that production AI Agent failures stem from fragmented context rather than model limits, detailing their ontology-driven four-layer architecture and three architecture iterations that transform context into governable, traceable semantic assets for stable delivery.
HOMURA introduces a reinforcement learning framework with KL regularization and dynamic syllable-ratio rewards to achieve precise length control in LLM translation for subtitles and dubbing, outperforming baselines on the Sand-Glass benchmark and deployed at scale on Bilibili.
Sierra released the Poppy protocol with major partners, enabling AI agents to authenticate via OAuth and operate with granular permissions via a /.well-known/poppy.json file, replacing the need for agents to mimic human behavior; the draft includes five permission levels, single-session continuity, and open Apache 2.0 licensing, though community debates remain about governance and revocation.
