Your map timed a jam that hasn't formed yet

transcript

5:40 PM. The Hebbal flyover in Bengaluru is flowing fine. Your map already planned for the jam there at 6:15.

Ananya needs the airport, 38 kilometres away. Her route crosses Hebbal. The naive ETA adds up every road's speed right now. Hebbal: 35 kilometres an hour. It misses the jam. It promises a time it can't keep. But she reaches Hebbal at 6:15. History says it crawls at 10 by then.

So walk the route in order, keeping a clock. Near roads? Live speeds. Further out, a blend of live and history. From 30 minutes out, history for that weekday and time. Hebbal, 35 minutes ahead, is mostly history. 10 kilometres an hour.

Why not live everywhere? Live data says little about 35 minutes from now. But history for that weekday and quarter-hour says a lot.

On top, a learned correction. DeepMind's graph networks look at stretches of connected roads. Up to 50 percent better real-time ETAs in some cities. Google reported over 97 percent of trips accurate.

At 5:40, the map made her a promise. She reaches the airport at 6:57. Within 3 minutes of it.

An ETA isn't today's traffic, added up. Time each road for when you'll reach it.

Full breakdown in the lecture.