This week: Baidu showed off Apollo World, an AI-built virtual city where its robotaxis can practice entire trips before they ever touch a real road. Every car and pedestrian in it makes its own decisions, and Baidu says one day inside equals years of road testing.

Elsewhere: How the simulation loop works, Baidu's new yardstick for judging world models, what we still don't know, and three more things that shipped this week.

The New Prime Signal

Teaching a car to drive itself has a slightly awkward problem: the rarest, scariest situations are exactly the ones you most need to practice, and you really don't want to wait for them to happen on a real street. Baidu's answer is to let the car rehearse inside a world made by another AI. At its AI Day event in Shenzhen on November 3, Baidu's autonomous driving unit unveiled Apollo World, which it calls an agent-driven, all-modality "world model" for physical AI. According to Reuters, the system is already being used to train and validate the fleet of Apollo Go, Baidu's robotaxi service.

What happened: A world model is an AI system that builds a working picture of an environment and predicts what happens next. Apollo World does this for city streets. In Baidu's announcement, the other vehicles and pedestrians in its virtual city aren't following a fixed script. Each one is an AI agent that makes decisions and reacts to what the robotaxi does. If the test car slows down or overtakes, the world around it responds differently. The system also generates what the car's sensors would see in real time, including camera images and lidar point clouds, and both update together when the simulated weather changes. The driving software decides whether to steer, speed up, or brake, sends that action back, and Apollo World updates the scene and produces the next moment. Baidu says a car can drive from any starting point to any destination in the virtual city, completing a full ride from pickup to drop-off.

Why it matters: Real-world testing is slow and expensive, and it hands you mostly boring miles. Baidu says Apollo World can generate up to 100,000 kilometers of simulated driving a day, which it puts at roughly three years of testing by a real car with a safety driver, Reuters reported. The company describes a loop: pull tricky scenarios out of real driving data, turn them into repeatable virtual tests, let the car practice, then send the updated software back onto real roads to check the result. That idea isn't unique to Baidu, but it is backed by a lot of real driving. Baidu says Apollo Go operates in 28 cities and has logged more than 350 million autonomous kilometers, over 240 million of them fully driverless. It also says the same technology could stretch to robots and other physical AI uses.

The fine print: Everything here comes from Baidu, and the mileage and fleet numbers are the company's own. The 100,000-kilometer figure is described as "up to," and "three years of testing" is Baidu's own comparison. We didn't find a technical paper, independent evaluation, or public demo alongside the launch, so there is no outside way yet to judge how realistic the virtual city actually is. Simulated kilometers are also not the same as real ones. A practice world is only as good as its physics and its agents, and a car that aces a simulation still has to prove itself on real streets, which is exactly why Baidu describes sending updated software back to real roads for checking.

The most interesting part might be the yardstick, not the city. Baidu proposed what it calls a "2+3" standard for world models. The two basics are causality, meaning the past should determine the future instead of the model stitching together nice-looking frames, and interactivity, meaning the world has to react when you poke it. The three advanced tests are physical accuracy, persistence (do objects and details drift over a long run?), and real-time speed. That is a refreshingly practical checklist at a time when "world model" can mean anything from a research system to a pretty video generator. A car doesn't care how cinematic the scenery is. It cares whether the cyclist on the right will still be there in a second. And honestly, if a robotaxi has to fail a few thousand unprotected left turns, we'd all prefer it did that in a city that doesn't exist.

Think of it as a driving simulator where every other driver is improvising, and the scenery is generated on the fly.

How it works: According to Baidu, each cycle runs in four steps: the world generates sensor views, the driving system perceives them and picks an action, the action goes back into the world, and the world updates every other agent and produces the next set of sensor views. Baidu says this closed loop isn't limited to one intersection or one stretch of road, but covers entire trips through different streets and traffic situations. Speed matters here. Baidu points out that a two-ton vehicle moving at city speeds has almost no room for error, and its safety decision cycle has to fit within one-thirtieth of a second, which is why it lists real-time performance as one of its three advanced tests.

Follow along: Apollo World is an internal tool, so there is nothing to download or try. If you read Chinese, or don't mind a browser translation, Baidu's announcement lays out the "2+3" standard in detail, and IT Home's write-up adds context on the testing loop. For the English version of the scale claims, the Reuters report has the numbers. The thing to watch next is whether Baidu publishes a paper or results that let outsiders check its own checklist against the system.