[Lily]

I was today years old when I first heard about Rivian.



Apparently, I'm very late to the party because their electric SUVs are super popular in the US, and every self-respecting tech bro drives one of these. (I might be exaggerating here a bit for good writing's sake, but you get the idea 😉).

I was reading an interview on Top Gear with Rivian's CEO, RJ Scaringe, and I was about to click on something else, when something he mentioned really surprised me:

"Previously, autonomous cars used a rules-based environment, in which humans tried to code how the world behaved (...) Now we’re picking up data from all the vehicles out there and using it to train a ‘large driving model’, which is learning to drive in the same way that a human or human brain would. It’s accretive, you can grow perception"

Oh boy. I never had much faith in purely rules-based systems because the real world is unpredictable. Rules change, people step into the road unexpectedly, a lamppost falls over, a dog suddenly decides to explore the street.

But a system that can learn from millions of real-world driving situations and adapt its behaviour accordingly? That feels like a very different proposition.

This is possibly a real game-changer in the self-driving race.

Can we really teach a machine to drive with something approaching human judgement? Could these systems eventually become safer than the humans currently behind the wheel?

[olduser]

I think, the article is misleading it is suggesting one or the other, rule based or accumulated knowledge, in reality, the rules of the road have to be in place first.
We are leaving out, the rules that the car uses to interpret what it 'sees.'
Most of the reported incidents with autonomous cars have involved errors in these rules.

Let's try to imagine a world with autonomous cars all using motoring rules.
Traffic will flow but I think slower, because the car has no choice but to follow every rule literally, every stop sign or line has to be obeyed.
Human drivers, tend to risk keeping moving, attempting to minimise the risk by judging if any other cars are going to arrive before the risker is out of danger.
The danger is, humans are not very good at estimating the speed of vehicles approaching head on, nor are we good at spotting motor bikes in one glance.

Would we want an autonomous car to develop this 'trick' ?
In practice it would probably see better than us, and calculate the time available rather than estimate if there is time to safely make the move.

Obviously, there many more examples of where human drivers bend the rules to keep moving, usually increasing the risks, we are very poor at estimating risk, what will AI make of these unless it is given a, very long, list of risks?

Up to date, AI only knows what humans already know, and have documented in digital form, it just finds it faster than we do.

I find it ironic, we talk of autonomous cars as being safer than humans because humans are not good at obeying motoring rules, then we suggest we should let them learn from humans!

[Santa]

So long as the computer can differentiate between good and bad judgment, it should be a good thing. The few times I have been in collisions, bad judgment has been the main cause. I am sure that this is the case in the vast majority of collisions.

It might be pulling out in front of an oncoming vehicle (look up 'saccades') or deciding to drive after drinking three pints of bitter. Computers don't drink and drive, don't get tired or distracted by passengers and try to go faster than is sensible for the prevailing conditions.

[Nick]

Bein a pedantic devils advocate for a second @olduser - it's probably relevant to recall that we were using motor vehicles on roads before there were rules, and the laws of the road came to be as a result of the way people drove and subsequent collisions (with other vehicles, people, and other things). I would also point out that this isn't the machines learning from us, as in, we're not teaching it anything. It's learning to recognise what is happening in human behaviour. Much like we might learn the behaviour of a group of animals - not so that we can become more like them, but so we can understand and be able to react and work around them - usually!

I think you're right in terms of judgement @Santa - these machines should be able to make a judgement based decision having "considered" significantly more factors than we are able to as humans - and much quicker to boot. I think this is a fascinating development, thanks for sharing @Lily

[NMNeil]

So the AI will learn from the driving habits of humans?

[Rolebama]

I had a debate with one of the AI Team from Loughborough University over 35yrs ago, which was simply caused by my asking which of the team of 5 working on the project would teach the AI? When we finished talking I was told that my argument had possibly set their work back by 10yrs, purely because the question could not be answered. (During the discussion I was told the reason there was an odd number working, was so that there would always be a majority when it came to disagreements.) So my question still remains, who considers themselves worldly and clever enough to program AI in this fashion? I would certainly enjoy a conversation with this perfect example of mankind.

[Santa]

@Rolebama

I think you are way behind as regards how AI works. If you ask an AI chat a question, it searches through a massive database to see how similar questions have been asked and answered. Most often, for straightforward questions, it gives the correct answer.

The problems arise when the question is poorly worded or complicated by multiple possibilities. I have also seen reports in which the chat engine tries to "please" the questioner with answers it computes will satisfy. It is this kind of response that might be the big problem with self-driving cars just now. One can only assume that the reference material will be carefully vetted.

[Nick]

I'm not sure I'd say @Rolebama is way behind - AI is a ridiculously huge concept, and there are definitely still models and platforms where data needs to be input for the platform to "learn" something. Usually it's exactly this - inputting data for it to learn from. These are often referred to as LLMs (Large Language Models). Sometimes these are connected to a live web search tool (such as Goole) but not always, and as such the data has to be periodically retrained and updated. However, as we saw in another post here recently, the way things are developing has led that specific system to almost be able to watch someone and learn to replicate the exact way to "physically" perform a task.

With regards to these self-driving cars, it's probably closer to the former, as in the platform will be fed a ton of data about driving habits, road uses, times of day, volumes of traffic, accident and congestion hotspots, car sizes, trends and patterns, and likely so much more - giving it the ability to make decisions and judgements based around all that data. I don't imagine there will be an observational aspect to this project.

[Lily]

@Rolebama I love the detail of the odd number team, reminds me of the "10th man rule" of World War Z: if nine people agree on something, it's the tenth person's job to argue the opposite, even if they personally agree with the group.

I think that now, more than ever, we need that 10th woman or man.

[Rolebama]

Twelve Angry Men springs to mind. Where 1 man disagreed with the other 11 in a jury trial, and gradually got them all to change their minds. Yes, I know it was a setup, but I have played Devil's Advocate, just to be bloody-minded and wipe the smiles off some smug faces. Then to tell them that I agree with what they originally advocated. I did this both in the Army and working for the Borough. Purely because they were working to a budget, and I do not think people are just statistics, yet from what I see and hear of AI, that seems to be an integral part of research.