If it's really end to end training where human driving is the primary training signal along with generative video prediction (that is what is being demoed) then no they are not the same ones at all.
Yes, that's the existing Karpathy solution on the perception side. It is trained against a large database of labelled video with human derived interpretation of 'objects', and then a rules & optimizer based policy planner is run with those features to drive the car.
Computational budget and power. They were already at the limits and gave up on the double-redundant so that each processor now does independent computation.
They are computationally limited now even after they do heavy sparsification and quantization.
The policy side likely consumed much less computation than anything neural.
E2E is far away from a production form, it's an entirely new approach and major architecture.
OK that's fine, but it also shows that they are budgeting microseconds.
Whatever the 'it' is it is years away in the future. They will always try to max out the computational ability of any hardware, you can always scale each net to a size with a given budget, but bigger nets have better performance and reliability. GPT-2 sucks at writing compared to GPT-4. Rats suck at quantum mechanics compared to humans.
I don't know anything about AI and E2E. But can the car move without being programmed? For example the system tells the car: stop in 100 m with deceleration rate of 10 miles/sec^2, turn 13 degree in the next 10 ft, turn on a curvature of 50 ft radius with speed of 25 miles/sec,...
I'm a little concerned that my reflexes may not be able to up to dealing with 25 miles/sec when it suddenly decides to take an unexpected exit short lane?
While not a direct correlation, it's a pretty easy road to travel, largely based upon what Musk and others at Tesla have said, that the entire reason for Dojo is to address computational performance and budget. In essence, Tesla dumping billions into the Dojo architecture proves out exactly what we're talking about here - that there are currently limits to both computational budget and associated costs at present - and that Dojo is being built to help address the current computational and cost limitations of the E2E NN.
This right here. With the CEO behaving like a decompensating mental patient, there's no way any company he leads is getting the very best talent, because that talent will go where it's more fully appreciated, rewarded, etc., without fear of having to deal with an overweening, preening, nasty, vindictive, lying blowhard.
Can you explain why Tesla has the market lead (beating BMW), best selling car (Model Y beating Toyota), and highest margin of any consumer EV being sold, all while having a CEO who's "overweening, preening, nasty, vindictive, and a lying blowhard"? Apparently he's doing it with substandard talent, because the very best talent isn't working there.
Because the manufacturing and mechanical engineering talent is really good, and there aren't wealthier competitors with tons of VC capital or gobs of cash flow trying to hire them away. And dealerships of course.
And this will be yet another excellent lesson on how the free market works. If what you say is true for the majority of buyers, the market will shift away from Tesla and towards the competition. Tesla will then either pivot and compete, or shrink in the market and eventually fail.
We are talking about runtime budget on the car, not offline compute when crunching numbers to weight the neural nets.
It would not surprise me in the least if Tesla uses it's own processors designed specifically for the vehicles. If Tesla is going to the trouble to build it's own processor network for AI NN central processing - it's not a stretch to see Tesla design it's own vehicle processors to complement specific needs at that layer.
? It did. Tesla hired Jim Keller to design its first in-car processor, and ever since then (all the way up through the current HW4) Tesla has been running their own chip design in the car.
I was of the impression machine learning is pervasive all throughout the code and without it nobody could possible hope to actually write self-drivable code.
The systems are better at detecting fast-moving objects, particularly out of front view, than many humans, particularly at night. I like using it for freeway driving at night for this reason.
Elon has personally stated that the high stock market valuation of Telsa is entirely dependent on FSD. He really really cares about his own money so he won't ever stop hyping it.
insideevs.com
insideevs.com
insideevs.com
With current management, pressure on margins and sales will likely lead to even more decontenting. That's the point of complaining about it before it happens.
And how is the complaining getting to current management? Surely you don't mean here on TMC. Are people submitting feedback to Tesla? Is Tesla holding focus groups to gauge reaction to potential changes? Typically, the free market is reactive, not proactive. A company makes a change to a product, and if sales decline beyond expected values, the company pivots and corrects the product (with or without feedback from the customer base - but usually with).
True, but only because Musk will have somebody on Mars by 2026
www.cnbc.com
