Google's WikiSkill gives AI agents a persistent memory of past mistakes to sharpen future performance
WikiSkill creates a wiki-like knowledge base for agents to store and reuse instructions for failures and successes, improving behavior over time.
319 candidate events; 60 shortlisted stories; 44 with article text available; 20 selected. Counts describe different stages; clustering, validation and refills can change the shortlist.
Active filters: none. Sort: Ranked.
All 20 stories, in ranked orderWikiSkill creates a wiki-like knowledge base for agents to store and reuse instructions for failures and successes, improving behavior over time.
Code-as-World converts real videos into executable MuJoCo physics programs to train physical reasoning, outperforming video-only models.
Tencent's Hy4 flagship model shows major performance improvements over its predecessor in benchmarks and early feedback.
OpenAI's Jalapeño is a purpose-built AI accelerator ASIC, not a repurposed GPU, with first benchmarks revealed at Hot Chips 2026.
Zhipu AI's GLM-5.3-Flash model, previously Ox Alpha, processed over 11 trillion tokens on OpenRouter in its first three days.
Recursive self-improvement is theoretically possible but practically difficult due to the time required for a single improvement loop.
Samsung is implementing MAC units within LPDDR5X chips to enable in-memory compute while preserving standard memory controller interfaces.
LAION released the Big Video Dataset with 10 million hours of footage, enabling models to outperform previous benchmarks by up to 2.1 percentage points.
Chinese AI firms are leading in video generation by leveraging vast short-video ecosystems, lower costs, and aggressive pricing.
AI Engineer Notebooks offer free, framework-free RAG/agents/evals on Colab to teach applied-LLM skills using raw APIs.
China has written its first national liquid-cooling standard as AI racks approach 1MW power density.
As embodied intelligence takes physical form amid widespread production-capacity expansion in the humanoid robot sector, it is becoming increasingly possible to envision a society…
Z.ai's new GLM-5.3 Flash model offers near-frontier capabilities at super low costs, intensifying global price competition.
Anthropic's Model Hardware Standard (MHS) cuts integration time for physical devices from weeks to hours.
Qwen3.5 on MI325X provides measured inference throughput and cost data for AMD's new AI accelerator.
Hugging Face and Pollen Robotics offer Microduck, a $399 open-source bipedal robot with a public sim-to-real training loop.
HarmonyOS has become the world's third-largest mobile operating system with over 1.35 billion devices in its ecosystem.
Baidu launched DuMateBench to evaluate AI agents on their ability to complete real-world tasks and deliver usable outputs.
PsiBot raised over $100 million to support embodied-world-model research and deployments in logistics and manufacturing.
NVIDIA Earth2Studio enables users to build custom batched ensemble weather forecasting workflows with atmospheric perturbations.
No stories match. Active filters.