AI
Intelligence Learns to Do the Chores, Build the Parts, and Practice in the Physical World
Intelligence spent Tuesday moving closer to everyday life, and the shift showed up in robots, chores, code, and engineering drawings at the same time. The common thread was usefulness measured in the physical world, with models judged by tasks completed, parts assembled, and designs produced. AGIBOT unveiled a robot brain that keeps learning after deployment, OpenAI published a benchmark built from real household work, Meituan shipped a giant agent model through a developer gateway, and a specialist engineering model matched frontier accuracy at a fraction of the compute cost.
AGIBOT unveiled GO 3 Preview in Shanghai on November 3, an embodied foundation model built on its World Reasoning Action Model architecture, and the model took first place rankings on three major embodied AI benchmarks, RoboDojo, RoboTwin, and RoboColiseum, according to the company announcement carried by AAP. The architecture unifies task understanding, future state prediction, and action generation, so a robot can look at a new object, anticipate what happens next, and shape its movements around that prediction. AGIBOT pairs the model with Genie Evolver, a post training system combining 100 physical robots with physics simulation and world models, and with Genie Master, an agent framework that turns daily experience into long term memory and improved action policies. A Genie G2 robot demonstrated the approach in a real convenience store, identifying causes of task failures and improving across complex manipulation tasks. AGIBOT plans full open source release of GO 3 Preview by the end of 2026.
OpenAI gave the industry a yardstick for the other half of everyday intelligence, the chores that fill a household calendar. LifeVal evaluates how effectively AI assistants execute real world personal administration, caregiving, and financial chores, with tasks constructed from time use surveys and feedback from a 3,995 person poll, according to Communeify summarizing the OpenAI announcement. GPT 6 Astra scored an overall average of 80.6 percent, compared with 49.3 percent recorded by GPT 5.2 nine months earlier, and reached 100 percent completion rates on 10 of the 16 primary task categories. Childcare coordination, food assistance paperwork, healthcare navigation, and personal finance management sit at the center of the benchmark. A benchmark built from how people actually spend time rewards systems that finish multi step work.
Meituan pushed agent capability to a larger scale through a channel developers already use. LongCat 2.5 arrived generally available on the Vercel AI Gateway on November 2, according to the November model tracker, with 1.6 trillion total parameters and about 48 billion active per token, a 1 million token context window, and pricing near 0.30 dollars per million input tokens and 1.20 dollars per million output tokens. The model couples visual comprehension with multi step tool execution for software builds, visual interface programming, and 3D design tasks, and it works immediately with coding agents such as Cursor, Codex, and Claude Code. A 1 million token context window lets an agent hold a large codebase, a design brief, and a long tool trail in view at once.
Deep Principle added an open weight pillar for science on the same day. Hydrogen 1.0 is a 753 billion parameter mixture of experts model optimized for advanced scientific computation and research, activating about 40 billion parameters per forward pass, with a 1 million token context window, according to Communeify. Checkpoints shipped under the MIT license with open weights hosted on Hugging Face and ModelScope for academic and commercial use. Open weights at this scale let a university lab, a startup, or an independent researcher inspect, adapt, and deploy the model on their own terms.
Skyfall AI showed what specialization can do against that scale. SKY 001 is a generative engineering model that synthesizes complete 3D mechanical CAD assemblies, pairing a Hierarchical Composer that decomposes assemblies into parametric part trees with a Construction World Model that predicts intermediate visual fabrication steps, according to Communeify. The model matches GPT 6 Astra on mechanical CAD geometric accuracy while cutting per design inference compute costs by approximately 150 times, a ratio that turns expensive engineering iteration into routine exploration. A mechanical engineer can now test more variants and iterate on assemblies at a pace that frontier general models price out of reach.
Anthropic rounded out the day with infrastructure for agents that keep working across long sessions. Message Threads entered public beta for the Claude Messages API and Claude Code, maintaining server side conversation state so clients send fresh messages plus the preceding message identifier rather than resending full transcript histories, according to Communeify. In requests spanning 500,000 to 1 million tokens, median first token arrival quickened by 35 percent, while median payload transfer sizes on long agent runs fell from 800 kilobytes to 1.8 kilobytes. Server side state removes much of that drag, which makes multi hour coding, research, and operations jobs materially smoother. Anthropic also updated its IPO filing during the same stretch, reporting third quarter revenue near 17 billion dollars, about twelvefold higher than a year earlier.
For readers, the day offers a practical way to judge the AI race going forward. Ask what a model finishes in the physical world, how it scores on chores drawn from real time use surveys, whether its weights stay open for inspection, and what each finished design or long session actually costs. Together they describe intelligence growing more capable, more measurable, and more affordable at once.
Quick answers
What is this story about?
Intelligence spent Tuesday moving closer to everyday life, and the shift showed up in robots, chores, code, and engineering drawings at the same time. The common thread was usefulness measured in the physical world, with models judged by tasks completed, parts assembled, and designs produced. AGIBOT unveiled a robot brain that keeps learning after deployment, OpenAI published a benchmark built from real household work, Meituan shipped a giant agent model through a developer gateway, and a specialist engineering model matched frontier accuracy at a fraction of the compute cost.
Why does this story matter?
For readers, the day offers a practical way to judge the AI race going forward. Ask what a model finishes in the physical world, how it scores on chores drawn from real time use surveys, whether its weights stay open for inspection, and what each finished design or long session actually costs. Together they describe intelligence growing more capable, more measurable, and more affordable at once.
Sources
New to crypto? Read the crypto glossary, browse frequent questions, read our story, or explore the story archive.