A new test checks if robot AI can predict real life, Malaysia seizes hidden AI chips, and two groups share ways to test AI helpers.
This week, researchers opened a test that checks whether robot AI can truly predict the real world. Malaysia also seized more hidden AI chips, and two groups shared new ways to test and guide AI helpers. The lesson for you is simple: when an AI tool shows you a result, check it against what really happened.
A reality test for robot AI
Some robots use a world model. That is an AI program that predicts what will happen next after the robot takes an action. It shows the guess as a short video. Think of it as the robot imagining a move before it makes it.
Researchers from Princeton University, the University of Texas at Austin, the University of Toronto, NVIDIA and AgiBot built a benchmark called WorldOracle, The Robot Report says. A benchmark is a standard test used to compare tools. WorldOracle gives a model the first frame of a real robot task and the list of actions. The model makes a predicted video. The platform then scores it against a recording of what truly happened. It opened to developers on October 16.
The team tested seven models, including DreamDojo, Cosmos 3, IRASim and GE-Sim 2.0. They found two main problems:
- The robot arm's motion was predicted fairly well. The objects it moved often drifted from reality.
- Models that did well on short stretches could get much worse over a full task.
In one case, the robot's grasp missed, yet the model showed a successful pick. Rankings also changed over time. IRASim scored below Ctrl-World, DreamDojo and Cosmos 3 over the first 30 seconds. Over the full sequence, it scored above all three.
What to watch: whether models get better at predicting failures, like drops and spills. WorldOracle counts a correctly predicted failure as a correct answer.
Hidden AI chips seized in Malaysia
Malaysian customs seized more than $12 million (49.5 million ringgit) in advanced AI chips, Tom's Hardware reported, citing Reuters. Officials found the chips inside 30 servers on October 15. The servers were in the free trade zone of Kuala Lumpur International Airport. They had been declared as ordinary electronic goods.
The airport's customs director, Aspalila Ag Tuah, said the scheme used Malaysia as a transit point. The goal was to avoid limits on where the goods could go next. Officials suspect the same group was behind a June 5 case. In that case, 72 server units worth about $13 million were found. She said the network is linked to seizures worth over $25 million across at least five cases.
Some background helps. The U.S. has export controls on advanced AI chips going to China. Export controls are rules that limit what can be shipped to certain places. Tom's Hardware says China is the most likely destination. Still, the chip type, final destination and buyer are not known. Malaysia, Singapore and Taiwan have set similar controls.
What to watch: more checks at busy shipping hubs, since AI chips are now treated as sensitive goods.
Measuring how AI agents run
An AI agent is an AI program that uses tools to finish a task, like reading a file or running a test. It often works in a loop. It calls a tool, checks the result, then calls another tool.
Tom's Hardware is building a test for how a computer's CPU handles this work. The CPU is the main chip that runs a computer's general tasks. The test looks at CPU and memory, not other AI chips. It checks how many agent sessions a chip can run at once while still responding quickly. The publisher says the test is a work in progress. The numbers are not ready yet.
What to watch: the first published results. They may show which parts of a computer matter most for agent work.
A coding team shares its checklists
MiniMax made its MCode Stack open source on October 28, MarkTechPost reported. Open source means anyone can see and use it. It uses the MIT license.
The stack has 15 skills. Each is a written set of steps or a checklist that an AI coding agent can follow. It holds no program code from MiniMax Code. The skills use a common format called SKILL.md, so most also work in other agents, including Claude Code.
One idea stands out. A skill called independent-review gives the reviewer only three things: the code change, the requirement and the test results. It leaves out the author's conversation. The reasoning is that a reviewer who sees the author's thinking may repeat the author's blind spots.
What ties these together
Each story is about checking claims against reality. Here are habits you can use with any AI tool:
- Compare what the AI shows you with what really happened.
- For long tasks, check the end result, not just the first steps.
- Ask for a fresh review that judges the result, not the story behind it.
Have a question about anything here? Email me at sales@appwt.com or call (888) 565-0171, and I will answer it.
Sources
- WorldOracle Tests Whether World Models Can Reliably Predict Real-World Executions, The Robot Report, 2026-10-28
- Malaysia seizes $12 million worth of advanced AI chips hidden inside 30 servers, Tom's Hardware, 2026-10-28
- Our first step at defining a CPU benchmark for agentic AI, Tom's Hardware, 2026-10-29
- MiniMax Open-Sources MCode Stack: 15 Agent Skills Its Coding Team Uses Daily, MarkTechPost, 2026-10-28
Tags
Frequently Asked Questions
What is a world model in robotics?
It is an AI program that predicts what will happen after a robot takes an action. It shows that guess as a video.
What did the WorldOracle test find?
Robot arm motion was predicted fairly well, but moved objects often drifted from reality. Models that did well on short stretches could get much worse over a full task.
Where were the AI chips found in Malaysia?
They were found inside 30 servers in the free trade zone of Kuala Lumpur International Airport on October 15. The servers were declared as ordinary electronic goods.
Are Tom's Hardware's AI agent test results final?
No. Tom's Hardware says the test is still a work in progress and its numbers need more work before they hold up.
Can MiniMax's skills work with other AI tools?
Yes, in many cases. The skills use the common SKILL.md format, so most work in other agents that support skills, including Claude Code.
Enjoyed this article?
Share it with your network
