Impossibility of L5 using any level of new hardware

[SergeyUser]

Still digesting the info from the Autonomy Day. Tesla explained how they train their cars. Well, I think unless Tesla changes how they teach their neural nets, with any future computer power, IMHO that level of L5 autonomy is impossible.

Why? Well, think about it. The way Tesla teach cars to drive is no split it into features, and train the NN to implement each feature. Some examples:

1. Traffic lights. Ok, let’s teach car to figure it out. One more NN.
2. Need to be able to interpret human signs? Ok, one more NN, feed lots of humans waving hands to train it.
3. Cars with bicycles? No need for bew NN, but people have to manually label data to mark ‘cars with bicycles’ as just cars.
4. Junk on the road? One more NN, feed lots of junk images to it.

So you get it 99%. Maybe 99.99%. And then you have to start to implement very weird, rare features, which maybe only 1 in a 1,000,000 person will ever encounter. Cost to teach the cars will become prohibitive.

So to go beyond that, someone has to figure out how to let the cars teach at a more abstract level on its own, basically feed traffic code + raw video feeds, without any labeling, triggers that feed corner cases back to Tesla for Humans to fugure out.

Now, there is no solution for that. At high level, car is constrained by a set of ‘features’, hard-coded by Tesla. This will stop the exponential improvement somewhere at 99.99%. This is not a Tesla-specific problem, of course. But this is why we are talking about 10+ years, until a more fundamental way for cars to learn by itsels how to drive is invented.

 

[SergeyUser]

So this is where in my opinion ‘exponential rate of improvement’ is not possible. It is limited by people defining features and working hard to produce specific training sets. It is too manual of a process to suddenly make cars smart, in a more general sence.

 

[woodisgood]

1. Car encounters edge case, becomes known to Tesla either by user taking over or AP deactivating.

2. Tesla queries fleet for as much video as possible of same edge case. (is this what you mean by working hard to assemble training set?)

3. Neural net trained either manually (human-intensive; this will become less common) or automatically (Dojo).

I’m not seeing the manual bottleneck? It sounds like they have the pieces in place to assemble training sets and train using them fairly automatically, or with much less manual input as they’ve had to put in up to this point.

 

[SergeyUser]

They extract cases based on ‘triggers’ defined by humans. Then they need to classify the data (I assume automatically) and then manually interpret it.

For many new features there will be new NN. Adding new NN is 100% manual. More rare cases will require more NNs. It is not like a human brain, but rather a fixed set of pre-defined ‘brains’. This is the bottleneck.

 

[commasign]

Doesn’t have to be perfect. Just several times better than the average human driver. The bar isn’t that high. Watching the demo video, I know people who drive worse than that.

 

[SergeyUser]

I agree, better than human is a low bar and achievable. I’m talking more about the limits beyond that.

 

[focher]

If it can pass a driving test, I guess we should let FSD drive.

 

[ewoodrick]

Still digesting the info from the Autonomy Day. Tesla explained how they train their cars. Well, I think unless Tesla changes how they teach their neural nets, with any future computer power, IMHO that level of L5 autonomy is impossible.

Why? Well, think about it. The way Tesla teach cars to drive is no split it into features, and train the NN to implement each feature. Some examples:

1. Traffic lights. Ok, let’s teach car to figure it out. One more NN.
2. Need to be able to interpret human signs? Ok, one more NN, feed lots of humans waving hands to train it.
3. Cars with bicycles? No need for bew NN, but people have to manually label data to mark ‘cars with bicycles’ as just cars.
4. Junk on the road? One more NN, feed lots of junk images to it.

So you get it 99%. Maybe 99.99%. And then you have to start to implement very weird, rare features, which maybe only 1 in a 1,000,000 person will ever encounter. Cost to teach the cars will become prohibitive.

So to go beyond that, someone has to figure out how to let the cars teach at a more abstract level on its own, basically feed traffic code + raw video feeds, without any labeling, triggers that feed corner cases back to Tesla for Humans to fugure out.

Now, there is no solution for that. At high level, car is constrained by a set of ‘features’, hard-coded by Tesla. This will stop the exponential improvement somewhere at 99.99%. This is not a Tesla-specific problem, of course. But this is why we are talking about 10+ years, until a more fundamental way for cars to learn by itsels how to drive is invented.

I didn't hear it the way that you did. What I heard was that they have a current set of features that they are looking for and they decide to add another set. They take the old set and add the training to it, and spit out a new set. Basically they are slowly adding recognition one step at a time as opposed to taking images and annotating everything to begin with on millions of images.

It's not as if they have to have a different NN for stop signs, left turn signs, right turn signs, caution signs, street signs, vertical traffic lights, horizontal traffic lights, cars, trucks, motorcycles, cars with bicycles, trucks carrying cars. It's one NN

And when the cars send triggers, they aren't expecting it to be 100% manual, they've got mechanisms that are automating that part as well.

 

[duanra]

My question is how to mixt safely robocars and human driven cars ?
If I refer to Musk reference with horses and cars. Mixing robocars and conventional cars seems as challenging as mixing horses and cars on a highway ! ;-)

 

[commasign]

If it can pass a driving test, I guess we should let FSD drive.

I dunno, there’s that one mean lady at the DMV.

 

[Paddy3101]

Still digesting the info from the Autonomy Day. Tesla explained how they train their cars. Well, I think unless Tesla changes how they teach their neural nets, with any future computer power, IMHO that level of L5 autonomy is impossible.

Why? Well, think about it. The way Tesla teach cars to drive is no split it into features, and train the NN to implement each feature. Some examples:

1. Traffic lights. Ok, let’s teach car to figure it out. One more NN.
2. Need to be able to interpret human signs? Ok, one more NN, feed lots of humans waving hands to train it.
3. Cars with bicycles? No need for bew NN, but people have to manually label data to mark ‘cars with bicycles’ as just cars.
4. Junk on the road? One more NN, feed lots of junk images to it.

So you get it 99%. Maybe 99.99%. And then you have to start to implement very weird, rare features, which maybe only 1 in a 1,000,000 person will ever encounter. Cost to teach the cars will become prohibitive.

So to go beyond that, someone has to figure out how to let the cars teach at a more abstract level on its own, basically feed traffic code + raw video feeds, without any labeling, triggers that feed corner cases back to Tesla for Humans to fugure out.

Now, there is no solution for that. At high level, car is constrained by a set of ‘features’, hard-coded by Tesla. This will stop the exponential improvement somewhere at 99.99%. This is not a Tesla-specific problem, of course. But this is why we are talking about 10+ years, until a more fundamental way for cars to learn by itsels how to drive is invented.

Your last two paragraphs have answered your own question, BUT the conclusion is that Humans figure less and less in the loop to 'figure it out'.

One of the fundamental things about AI using Neural Networks, is that they aren't taught to 'remember' things, they are taught to 'understand' things. They can (and in-fact they always do) provide and answer by generalization, they will give an answer based on the rules, they have learned.

You start off training the NNs with labeled data to get them to a point where they can do the basic functions, and have a base level of capability.

So labeled data is a bit like learning your times-tables, you get told the question, and then given the answer over and over again, until you get it. Then you can apply that to unusual examples you haven't seen before.

Then you can move on to 'Semi Guided Learning' whereby you direct the NN to the new things you want it to learn. The child has grown up a little, they are at senior school, and some of what then learn is from Mentoring. (i.e. They watch and follow a mentor, and learn from them - In otherwords the AI is watching and learning constantly from all the human mentors which are driving around)

When this is done (feature complete) then the AI can do all the functions it needs, and has been trained to do it competently (but not necessarily expertly).

You also build into the architecture, a feedback loop (which they almost certainly have if you listen carefully). (A Conscience if you like, a knowledge of what is good and bad) That feedback loop is autonomous, without needing manual teaching. You have a capability for the AI to judge for itself what are good and bad outcomes. At this point you have a single 'consciousness' which is learning to refine it's behavior based on the Millions of instances that are driving around, and getting feedback, both positive and negative outcomes.

There is still the Mentoring by the human drivers, but at some point, the pupil exceeds the knowledge/capability of the mentors, and the AI is just learning, unguided from it's own experiences in the real world. This is the march of the 9s. This is where the more data/results you throw at the AI the more refined it's knowledge/perception of the world becomes.

So when you have millions of AI drivers, driving 100's of millions of miles per week (conservatively), with a much better set of sensors than a human, 'COLLECTIVELY' learning as a single entity/mind, then it's going to get better, pretty quick, once the basic training by Humans is out of the loop.

This isn't 'Magic' or particularly new technology, what is New, is that the processing power of computers has reached the stage (predicted 30+ years ago) where they are fast enough to do the calculations and learning at a practical/useful rate at a reasonable cost.

If you want to read up on some of this, lookup 'Mind Children by Hans Moravec' which was published in 1988, which more or less gave the date of 2020 when this the capability would be available, practical, at an affordable cost, whereby the mass proliferation of individuals needed to 'experience' the world in order to learn would be achieved. (Based on Moore's Law)
Hans Moravec - Wikipedia

 

[odddiv]

Every time a human takes over, or does something that the current NN doesn't accurately predict when driving without FSD, that's a "trigger". Today those triggers are inspected and labelled by humans and fed back into the NN. My understanding from yesterday is that these triggers - where the NN has a discrepancy - are getting lower and lower in the shadow environment. Which is not the same as what we currently experience when driving - they haven't turned on all of that learning yet because they want to benchmark it in the real world, without actually impacting safety.

Dojo, I expect, will begin eliminating the human labeling from the feedback loop - as mentioned in the above post. The NN is getting "smart" enough that it can handle blending outliers in on it's own.

 

[ItsNotAboutTheMoney]

Why do people always assume that an autonomous vehicle has to be perfect? It doesn't. It just has to be very good in almost all circumstances and then _safe_ in others.

 

[banned-66611]

If it only has to be better than an average human driver then all the extremely rare corner cases are not so important.

 

[Paddy3101]

If it only has to be better than an average human driver then all the extremely rare corner cases are not so important.

From a 'Regulatory' point of view, the definition of 'better' is probably quite narrow. Where 'Better' means 'Safer'. That is being measured using crashes/mile stats.

After that, it's all a matter of marketing/personal opinion. Does it achieve what I was it to do?

For me, it doesn't need to drive how I drive (My wife doesn't drive like me at all), just needs to get me where I want to go, reasonably efficiently in terms if cost and time, without fuss.

 

[duanra]

Ok, better than a human driver in enough in theory right ? Now let's consider the psychological side of it.

Let's say you're driving along in your aotonomous car (no driving wheel).
And you see the accident coming, but there is nothing you can do...I bet it will take you a while before you step in one of these car...

 

[StellarRat]

I don't expect true L5 anytime soon. I do think full auto drives on well maintained and established roads to come though. Personally, I'd have a lot of trust issues with self drive until 1,000,000s of miles of trouble free drives were logged. Also, you have to realize that even a self driving car can't avoid all accidents. Some things are unavoidable.

 

[WileyTheMan]

I predict Tesla will get there, but Elon's time frame of his fully autonomous robo-taxis will be off by a few years, maybe 5-10.

 

[Darscot]

Ok, better than a human driver in enough in theory right ? Now let's consider the psychological side of it.

Let's say you're driving along in your aotonomous car (no driving wheel).
And you see the accident coming, but there is nothing you can do...I bet it will take you a while before you step in one of these car...

You mean like riding a bus, taxi, or with my wife on a daily basis I don't really find being a passenger that big a deal.

 

[ezevphl]

I had an Autopilot trial in August, I would take over fairly frequently. Now, half a year later I have AP and I almost never need to take over in the same situation. The NN learned from all the input and got much much better.
Is it there, no. Will it be there in a year? maybe not quite. I think it will get there and on that 99.99 percent will be much better than a typical driver.