That works for games in which the complete rules can be described. It doesn’t work for “games” (like driving a vehicle) for which the complete rules cannot be described. That’s why babies (and puppies) might misbehave in a restaurant the first few years until they observe the implicit rules of the place (“road”). And, the simulated toy restaurant that a child has DOESN’T help one bit in learning those implicit rules.
1 Like
Cops and robbers? Hide and seek? Fishing? Dating?
The Captain
1 Like
Exactly! You don’t see AI doing those things very well via simulation learning.
I mentioned this to my wife and she replied that having one stubborn @s$ in her life was plenty.
3 Likes
The Captain
2 Likes
One complication in hitting that target, though, is that it’s a little uncertain what it means to be “better than the average driver.”
As we discussed upthread, a disproportionately large number of accidents (especially serious accidents) are caused by teenagers and by drunk people. Or drunk teenagers, since they’re not exclusive. But most drivers at any given time - which would certainly include the median driver - are neither teenagers nor drunk. Negative outcomes, again especially fatalities, are highly skewed towards a very small subset of the driving population. Which means it is very possible to have an AI system that’s better than the average driver, but worse - perhaps significantly worse - than the average sober adult driver.
In that situation, the typical driver (a sober adult) would be making things more dangerous if they activated their AI, even if the AI was better than the average driver. The machine would never be drunk or teenaged; but since it’s a worse driver than a sober adult driver, that driver would be better off without the machine.^^
That’s why being better than average, even significantly better than average, is probably not going to be good enough. I don’t think regulators will allow too many situations where someone is seriously hurt in an AI-driven car that could have been avoided by a reasonably good driver. IOW, the AI isn’t going to have to be significantly better than the average driver, but instead will have to be significantly better than the average adult, sober, and attentive driver. At least before it gets regulatory approval.
Albaby
^^Note that if everyone were activating their AI’s, then it’s possible even that median driver would be better off - because even though they are better than the AI, the safety from not having teenagers or drunk people trying to drive themselves might compensate. But since AI driving will likely be rolled out piecemeal, the regulatory safety analysis will almost certainly require the AI to be safer for the occupants of the AI car in isolation.
1 Like
The “machine learning” that is going on now is not very different. If you look at how Alexnet won the Imagenet competition back in 2012, the approaches being used today are very much the same…many layers in a convolution neural net (CNN)…thus the term DNN (Deep neural networks). The industry has just built more optimized software and more optimized hardware to run bigger and bigger networks solving bigger and bigger problems faster. Most of the change is adding some new and more optimized layers in the models. The basic principles are even the same as what was used to read zip codes (see MNIST) back in the 1990s.
Maybe this example will help.
In the game of Go each side can only do one thing…place a game piece. Each side can only do the same things the other side can do, thus it can learn some new strategy based on how it won or lost or gained/lost some squares. This learning is how it then makes that strategy part of its tool bag for future games.
In a driving simulator the simulator is basically generating the world as a 3D game and controlling (and randomizing) all the actions, while an autonomous car has its driving software running inside the “game” trying to safely get to its destination. The 3D game simulation does not learn from the actions of the autonomous car since it is not trying to copy what the car does.
I suppose that you could make a simulator observe and see how a car got out of a sticky situation and make the game harder by recreating that situation and blocking the safe path. But this does not allow the simulator to learn new edge cases, such as figuring out that it can cause a power outage that turns off all the traffic lights. The programmers would have needed to think ahead and put that feature in. Same thing for lanes closed for maintenance, or icy roads, or foggy conditions, or mud on cameras, or sink holes or any of hundreds of other edge cases – the simulator has no way to learn how to do those things from the other “player” in the game.
If you disagree, please find a link to a description of any simulator that does learn new things (edge cases that must be tested) from the autonomous car.
Mike
4 Likes
Seriously, what part of “we don’t know and I think we’ll be surprised by how fast progress is and how it happens” are you not understanding? Tediously explaining how you think some practices you think you understand won’t lead anywhere, and how you don’t see any other ways to do things are not useful or enlightening observations.
We don’t know. I don’t know. You don’t know. You’re not going to suddenly know any more by discussing it with me. Right now, things are in flux and even the researchers in the middle of trying new things don’t have time to read all the papers being published on other new things.
This is what a revolution in scientific understanding looks like. Go read Kuhn. It’s good stuff. I just wish I were young and starting out again rather than old and watching from the sidelines.
-IGU-
1 Like
Are you able to answer this question on why cross validation can be a nonsense idea, in your view?
I didn’t see an answer in your reply.
2 Likes
I think you are trying to convince yourself that there is something magical (possibly) soon to be discovered in machine learning.
But think about this. The Tesla HW3 is basically a very fast and power efficient DNN accelerator. And HW4 is supposedly the same thing, just 2-4x faster to be able to process higher resolution cameras.
Possibly the best autonomous driving HW available…but it is not going to magically solve all the edge cases, IMO.
Mike
2 Likes
From Wikipedia:
Kuhn did not see scientific theory as proceeding linearly from an objective, unbiased accumulation of all available data, but rather as paradigm-driven:
One I like for its simplicity is a law of motion:
Aristotle: things move when you push them
Newton: thing keep moving until you stop them
Is ICE to BEV a paradigm shift?
is ICE to BEV plus AI a paradigm shift?
The Captain
1 Like
Should we add:
Is the shift from the legacy business plan of “you got what you bought and you don’t get any upgrades until you buy a new car” to the Tesla plan of “continuous OTA upgrades to keep even those cars sold a couple years ago, up to “new” specs”… A paradigm shift?
ralph
6 Likes
Interesting question! I suppose it depends on how paradigm is defined. Why would you want to give away things for free? In business because giving it away for free has greater benefits than not doing so.
Classical economics dealt mainly with the Law of Diminishing returns but information technology (IT) changed all that. The IT economy is ruled by the Law of Increasing Returns, once you create a software product, making and distributing copies of it is practically free of cost. OTA you don’t even have CDs and shipping costs. The recipient is paying for the Internet connection!
If you develop software for your new products you might as well make it available to your loyal customers since OTA already pays for itself by reducing physical recalls.
I would call it more consequence than cause. You make more money by selling more cars instead of selling more expensive cars. A competitive advantage.
The Captain
3 Likes
That’s only partially true. If a customer parks their car where they have a WiFi connection it is true. But for people that do not they get the OTA updates via cellular data and I think Tesla is paying for this bandwidth.
Mike
Just saw this video. Bill Gates takes a drive in an autonomous car in London. Very impressive tech.
3 Likes
Because it’s fundamental only to the way we do things now. It’s not necessary at all. What’s necessary is that the end result of the machine learning is something that works better for us than what we had before. Job done.
Let’s say we get to the point where we can solve the self-driving problem by simply having a humanoid robot get in any car and drive it without any specialized training. And let’s say one day the manufacturer, who has been selling these general purpose robots for a while, says they can now be used for driving and are insured for liability by the manufacturer. All it requires is paying an extra $50/month.
There’s no cross validation involved. Because the thing wasn’t ever trained to do that particular task. Rather it ingested all the experience of all of humanity that is in a form available for it to ingest. Then it figured out how to apply its “experience” it to the task at hand. My understanding is that things similar to this are being done now. Generalized training is working better for specialized tasks than specialized training.
-IGU-
2 Likes
The Captain
When you buy an iPhone you usually buy some other services a la carte, and continuing. AppleCare is one, iCloud is another, music subscriptions and others come quickly to mind. This is not particularly new, Ma Bell used to sell you a phone and line, but if you wanted “an extension” you paid more, even though it cost them almost nothing to provide it. Want Caller ID? Call Back? All those were an up charge.
It hasn’t worked that way in automotive because it was expensive to retrofit the beast, but when software doo-dads can change for the price of a download, I expect more of it. I also expect they will continue to provide free updates for certain things, just as Apple does as it upgrades its core system.
I won’t like it, but I won’t be surprised,
4 Likes
Everybody likes it. In fact everybody loves it. Buying something physical and having it get better over time, with new capabilities and features you don’t have to pay for is magical.
My many Apple devices and my Teslas are all better than when I got them. That regular software updates make my Teslas materially better every few months is one of the best features of the cars. The fact that I can buy even more stuff for them is an extra bonus, not a detriment.
As a Tesla paid upgrade example, since I got my Model Y, at the end of 2021, it offers me the choice (in the Tesla app, under Upgrades → Software Upgrades) of paying $2000 for the “Acceleration Boost”. It is described as “Improved 0-60 mph time, 4.8s to 4.2s”. Although I have no particular desire or need for this, I like the fact that it is there. Maybe if they put it on sale some day, I’ll splurge.
The actual advantage of this upgrade (as I know from my Model S), is the big improvement in 40-60mph time. That enables passing easily on twisty roads with limited legal opportunities.
-IGU-
1 Like
As a road warrior I attest that instant acceleration can be a life saver.
Add-ons are one thing, updates another, and the twin sometimes meet. (With apologies and thanks to Rudyard Kipling).
The Captain
1 Like