The current generation of neural networks already seem to do pattern recognition better than humans. And that is all that is needed to press the brake in a dangerous situation. All that is needed is to train them on enough examples of different situations that occur on the road.
As I said, neural nets are already good at pattern recognition. That is all that is needed to recognise you ran over someone.
Humans are still better at long reasoning sessions. Aka "thinking something through". But that is not involved in recognizing you ran over someone.
So this is not an example of pattern matching being insufficient to avoid danger.
There’s a lot of cars to replace. Capturing 10% of yearly car sales, worldwide, is already a very tall order.
Capturing 10% of all cars on the road would require this level of gargantuan sales for many years, and that’s just so 10% of cars in the world are your brand, not so 10% of rides are on cars owned by the fraction of people who bought your brand and decided to let it drive strangers while they’re not using it.
Moreover, beyond the safety concerns there’s policy concerns— knowing how to safely drive in Indonesia doesn’t mean being allowed to drive in Indonesia. Resource constraints— where is all the lithium going to come from, and how will the people around the world feel about this?
And there already are alternatives beyond the car industry. Public transit, for example. Demand for not driving or not owning a car can’t only be safely fulfilled with robotaxis.
That is 1% of all rides taking place in Tesla robotaxis, not 10%.
Also it is about rides, not about cars. Less than 1% of cars being Tesla robotaxis is enough for 1% of rides to be Tesla robotaxi rides. A robotaxi will do multiple times more rides per day that the average car.
The only real way this would work is if you build a pipeline where you have a simulated world with realistic physics, rendered through some pipeline to represent camera images (in relatively high fidelity) and/or lidar outputs, and teach the network to predict the evolution of these scenes, and do this with a shitload of random data
This is probably orders of magnitude more expensive to do than training GPT models, since for vision, you would have to render it with ratyracing.
Human drivers drive with EXCEPTIONALLY more "software" under the hood. We simulate the world around us using learned rules across multiple domains. For example, if someone has never driven at night, but understand that the cars have tail lights that are red, we can deduce that pairs of red lights are cars on the road without ever being subjected to the visual of driving at night. Its this kind of processing that lets us be safe.
In the current form, ML models are way subpar to this. A superhuman driving agent should be able to drive through a construction site with cones and debris, or a busy parking lot, or a grass field full of logs. No model out there can do this. This is why you absolutely need simulations if you want to do it traditional way. Comma AI has gained quite a bit of success in training in a simulated environment that simulates deviations from straight line driving on real world video.
we can deduce that pairs of red lights are cars on the road
Even if you had a simulation set up for training, and enough compute on the vehicle to run something like MuZero real time inference for self driving, you would still run into the problem of humans setting up all the possible driving scenarios, which will likely leave some out and the model will never learn the right actions for those - meanwhile, to any human, the action would be very obvious. Over time you could probably get close enough to a very very small error rate, but you would still be hesitant to trust the system.
A true superhuman driving agent would most likely be able to take a picture of a scene, and then give a prediction of evolution of object position in 3 dimensions for a given time window, and give a confidence score on that. For example, if the single image is from few on a highway, it should be able to predict that cars are in fact moving, because the chance of cars standing still on a highway is very low.
And to train a model like that, you most likely need a base model that can "understand" physics.
Additional images from the past or the future would generate prediction that is more accurate, and also improve the model self assessed confidence. And then you would have some heuristic algorithm like MCTS based on confidence levels on the best course of action.