Day 21 - Meetup: data and autonomous driving

At the 7 August meetup at Casino del Masnou we talked about autonomous driving and data. It was not a product pitch: it was a table talk on which level we are really at, and why training these systems burns absurd amounts of information.

Which levels exist (and where we are)

The SAE scale runs from 0 to 5. Useful summary:

• Level 0: the human does everything.
• Level 1 and 2: assistance (lane keep, adaptive cruise). The driver stays responsible. That is where most street cars live, including Tesla Autopilot / Full Self-Driving in daily use: you cannot leave the wheel for long or the car warns and eventually brakes.
• Level 3: the car can drive in bounded conditions (for example highway, up to a speed limit). Mercedes reached that with Drive Pilot on the S-Class; liability in that mode sits with the maker. The package is expensive (duplicated sensors for safety) and commercial uptake has stalled.
• Level 4: high automation in a defined domain. Waymo-style robotaxis in some US cities run with no driver on board.
• Level 5: anywhere, any condition. Still not a consumer product.

For the average buyer today, the realistic ceiling remains level 2. Anyone aiming at 4 needs another scale of sensors and compute.

Two sensor philosophies

Two clear approaches:

1. Mostly cameras (Tesla-style): cheaper, but fragile in fog, heavy rain, low sun, or night.
2. Sensor fusion (cameras + lidar + radar, like Waymo): more robust and far more expensive to build and keep in sync.

A lidar “sees” hundreds of metres; several lidars, radars, and a dozen cameras generate terabytes per day per car. Syncing those signals in real time is a computer on wheels, not an add-on.

Data, rules, and AI

Fleets already on the road (especially Tesla) pile up miles every day; that feeds training. Whoever does not keep that telemetry at scale cannot compete on driving models.

Mercedes built much of level 3 with hand-written rules (hundreds of thousands of lines of C) before generative AI was everyday. Parts of the industry are pausing level 3, staying on 2, and looking at 4 with neural nets: throw away rule code and train again. Training costs city-scale data-centre power; putting the model in the car remains the bottleneck.

Close

If you care about the conversation (data, AI, or how we run KM0), come to the next meetup. Between meetups, Cloud and Mail stay in the EU.