Yes, I believe these charts are about the compute cost of model training (versus running a deployed model in the vehicle which would be something that “goes on” (happens) during driving).
But training is required to have a model that can drive the car.
And compute cost (number of calculations per second) to train and run the model while driving (also called inference) both would increase with increasing model size (parameter count).
Key point:
If the compute is technologically or cost prohibitive for training, then no AI driving happens.
(that’s why people build these charts, to try to understand what is possible, or not, as model size grows)
Well, I think they are about you, me and everyone.
These statistics estimate the safety metrics of a random (“average”) rider or driver, as the case may be.
In the same way a clinical trial, which I may personally not participate in, estimates the safety and efficacy of a drug that you, me or anyone might need to take.
Sure I would. They’re not absurd, and they have nothing to do with anything you can see when you use FSD.
FSD is an impressive piece of technology and it’s almost like magic when you watch it drive. But the critical bottleneck to autonomy (removing the driver from the car) is about the “march of the nines” failure rate - not how well the car drives in any single trip.
That’s right - it’s about sample size. We tend to talk about how bad humans are at driving, but they’re actually really amazing at it. Nationally, the rate of accidents is about one per 700,000 miles (there’s obviously lots of different estimates, but that’s the one I’ve seen most often).
That’s so small a failure rate that you can’t assess whether an AI driver is coming close to that level through your own personal observations of a single car, except over several years. You’ll only drive about 10-12K miles in a given year. In that year, you’d never be able to know whether your car’s AI driver had a failure rate of 1:700K (comparable to a typical human) or 1:70K (10x more dangerous than a typical human). Because the human failure rate is so incredibly low, a car can have a failure rate that’s still an order of magnitude higher and still be unlikely to show up in the sample size of a single user, even over a year or two.
So it’s foolhardy to think you can learn whether FSD is close to “true” autonomy (for any commercially viable definition of “true”) just by watching the system drive for a few weeks or months or even a year or two. You won’t get a large enough sample. Even if it never disengages over that year or two, that’s not enough for you to know that it’s close to human safety; even if it does disengage once during that time frame, it might be a fluke.
You can’t learn whether FSD is close to self-driving by experiencing FSD. That’s just falling into the trap of thinking that AI’s have the same kind of correlation between quality and error rates that humans exhibit.
That’s not an insult, it’s a basic question about ability to engage in a two-way conversation.
And you are not able (or willing) to answer an honest, good faith question posed to you - here in this thread from me (and from others) and in other threads.
Once again you state and assume someone is all wrong, but you can’t be bothered to explain why.
This happens repeatedly.
No, this is flat wrong. It is accurate.
AI is built from mathematical formulas (which can be very complex) with parameters that make a mathematical function. Parameter values are identified during training.
This is foundational.
Many introductory books that have some math will cover this.
Nope, not an oversimplification. Functions need not be simple.
Nope, not ignoring learning/training. A trained model is just a (very complex) function with specific parameter values (identified during training).
Still a function.
Made of formulas.
Having parameters with values set during training.
Note how the loss function is written in a nice, tidy formula.
Nope, no intent to conflate.
As I wrote, I am just talking about model selection, a basic machine learning concept.
You seem, as best I can tell, to be thinking about a much more abstract concept like “use AI to architect/design/formulate a bunch of candidate models and then pick the best one” like you mentioned upthread. But as I said upthread, I am not talking about anything that abstract and complicated.
See how I answered your question (again)?
I wish you would do us the same courtesy and answer the questions posed to you.
Instead, you send more questions.
That’s not a two-way discussion, it’s more one-way. I am saying this with the absolute most positive and good faith intention.
I can see more clearly now that you really don’t understand my statements about machine learning.
This statement is a very basic way to state the process of model selection:
I don’t know how else to word it succinctly.
There’s nothing fancy or complex in what I am writing, I’m using the basic language of machine learning.
Again, all that question shows is that you haven’t understood, or perhaps even read, Sutton’s paper on The Bitter Lesson, because if you had, you wouldn’t be asking that question.
Which is what I said up-thread, but to which you didn’t respond.
Sutton writes of “leveraging human knowledge” by incorporating human knowledge into models. Like hard-coding a rules-based decision into a model or a system constraint (always stop on red) or something similar.
But your wording, ”the best architecture is NOT one that mimics the human mind’s approach,” is not the precise message I see in Sutton.
This stood out to me as weird (as I wrote upthread) because everyone is using neural networks which do indeed ”mimic the human mind’s approach.”
But I do think the second clause you wrote, ”but one that is general purpose…” is on target.
Sutton is at the leading edge, favors general models, general learning approaches, I get it.
I favor general models at the leading edge.
Most people with advanced training and experience in ML will agree on benefits of general models (I haven’t done a poll, but I am familiar with this world.)
So I asked a basic question:
What is your answer?
Are you willing to engage a question about your statement?
It’s not the end of the world to engage someone with their question.
OK, I so used “human mind’s approach” instead of “leveraging human knowledge.” If you had trouble with this difference in wording, I’m sorry, but you could have expressed a question about that difference directly, instead of insulting me (“Are you able to answer a question posed to you?”). And when I deferred to Sutton again as an answer, that also should have been good enough. At any rate, an honest, direct question that wasn’t phrased as a gotcha would have had the simple answer that I was using a different set of words to describe the same thing.
And after all, Sutton himself phrases the concept in different ways. Sometimes he says things like “leveraging human knowledge” but other times uses “human-centric approach,” or summarizing as: “building in how we think we think does not work in the long run.” Do you also question Sutton in not being precise on the differences between “knowledge” and “approach” and “how we think?” Because those are different at the same of precision level you claim to be striving for.
So, maybe you want to ask Sutton about why he talks about “knowledge,” which typically refers to facts or information versus “how we think” or “approach,” both of which involve the more active process of solving problems or creation of new ideas.
It was, and remains, a gotcha question posed not to gain clarification or a deeper understanding, but instead, as your insulting follow-up question makes clear, designed to put forth your own agenda. I’m not going to play your games.
You drilled me on this over and over, even after I clearly indicated that I was summarizing Sutton in my own words, and I even tried to clear your “confusion” by pointing you back to Sutton, re-quoting a different portion of Stutton - only to have you respond: “I didn’t say anything about Sutton.”
Remember, this is just “an honest, good faith question” about ability to follow simple logic. I had thought you were playing a gotcha game but you’ve effectively denied that now, so this is the only other explanation I can think of.
And this happened repeatedly: I count 6 posts from you on it. So it’s wrong to characterize my response now as “an over-reaction.”
I don’t think there is any good pure play on self driving. Tesla robotaxi is buried in a car business that is almost surely going to become low-to-mid-margin like all other car businesses. Waymo is buried in Google, a large advertising company. And things like Mobileye are a crapshoot because they don’t really “do” the business, they license “stuff”, and while that could become a great success, it isn’t likely to be a blockbuster.
Nobody ever expected autonomy to be based on v2v technology! What we did expect is that autonomous vehicles would use v2v technology as a tool to be more effective and more efficient.
I’m not smorgasbord1, but I’ll take a stab at this question.
Deep Neural Nets do, in some ways mimic the human brain. But in many ways they do not. Let me go through a few points.
DNNs, during training, use what is called back propagation to correct errors as the huge data set is repeatedly fed into it. Each error provides a small correction the the weights of the model based on training parameters. Pretty much each correction is the same. Humans do no such thing. One big obvious error to a human can cause a large improvement in getting the correct answer, such as that big cat isn’t a lion it is a leopard because it has spots. Just one instance of this is generally enough for a human to fully learn this without hundreds or thousands of repetitions.
During training DNNs use a technique called dropout, where randomly a large quantity of the learned weights are zeroed out and training continues with a degraded inference until it relearns those weights. This helps DNNs because they sometimes learn irrelevant stuff (i.e. leading to hallucinations) and after the dropout it is relearned if it was consistent and important. This would be like a third grader who previously perfectly learned the multiplication tables, to forget 10% of it, then relearning it again in fourth grade.
Using neural nets there is a distinct training phase. When using the trained DNN model we call this inference. During inference if the model does something incorrectly, minor or major, it will continue to make the same mistake(s) over and over never understanding or realizing its errors (unless there is some other external system checking the results) The only way to “fix” this is by another technique called reinforcement learning which is really back to a different type of training phase creating new weights for the model. This would be called a learning disability in a human.
One could go on and on with these differences which have improved how ARTIFICIAL neural nets have functioned but have diverged from the idea of how biological neural nets are thought to work.
Elon appears to be doing his best to kill Tesla’s personal car business. On the recently introduced long-wheelbase Model YL, with more roomy third-row seats, he said Tesla might not make them for the US market because of autonomy:
Which is a silly reason. It’s not like people don’t choose vehicles with more seats and/or more rear seat room for taxis/limos today.
For good or bad, Musk is betting Tesla’s future on autonomy, including robotaxis. Whether that’s a good or bad play depends on your outlook on Tesla’s success in that area.
I never understood that expectation. I actually interviewed at a V2V technology company several years back, and they couldn’t provide a good answer to my question on how this helps given that there will be non-V2V capable cars on the roads for decades to come.
Note that V2V is cars talking directly to each other. Most tech today revolves around V2G and G2V, where “G” is the grid, typically a cloud server. Cars reporting potholes, accidents, road closures, etc. and then the cloud telling other cars about them. The most common usage is, of course, traffic reporting used for optimized navigation.
I suppose there still are roads in the US on which there is no cellular reception, but those are also roads where V2V wouldn’t help much.
We could say that AI NN mimics the human brain in structure (having threshold-like functions, neurons, that are connected in a network), and to some degree in function, in that there is transmission of information among the neurons through the network.
But in other functions, such as learning, as you illustrate, the way AI NNs learn is probably mostly a human construction that differs from our best understanding of human learning.