Tesla: Artificial Intelligence and Autopilot
tesla.com
tesla.com
They're using transformers to persist occluded objects which is very cool.
Path planning is done almost entirely without neural networks. The do have an optimization routine which runs on the car with cost functions for comfort, distance, etc.
They have HD maps but the data is stored in the vision network's latent space. A joke was made about the vision network overfitting and storing a representation of the whole world, ie an HD map embedded in the network weights.
Also a comment was made about detecting pillars, seems that's on top of mind for the team.
I have never really seen a system integrated like that? Are there other examples of people doing this? They seem to have designed some kind of sled that connect to power and liquid cooling system.
Can somebody with some expertise in large scale training compare this to state of the art?
In Q+A they said the Dojo stuff was an aspirational vision, not a production-ready design
Elon is a self parody at this point.
How did you get that interpretation from Elon saying they're aiming to have a prototype next year?
It's a recruiting event. Presumably in part to recruit people to work on this new project.
And Dojo is clearly advanced state, as they show the hardware and showed how far they are. They haven't built the full rack of course.
> Elon is a self parody at this point.
I think people who act like everything he says is some made up nonsense are a parody of themselves if anything.
There's a much higher chance there's a consumer fraud case on FSD than ten of these robots are ever sold
- Communication speed is one of the biggest bottlenecks to large models, so their bandwidth of 4TBps is very smart.
- They claim a 1.3x perf/watt improvement, which is not really that great for ASICs compared to GPUs. Perf/watt is probably the most important number in datacenters.
- They only use SRAM, no DRAM. This is a huge mistake, which limits their model size. You can only fit a ~10GB model inside a single tile, versus 80GB models for a single A100 GPU.
- Software / compiler stack is as or more important than the hardware itself, because it dictates how much real performance you can squeeze out of the chips. I think Tesla will need to heavily focus on this area before getting anywhere close to real-world GPU performance.
This seems pretty interesting. They must have had a reason for this approach. Maybe it doesn't matter as much for their usecase.
They clearly had the option of adding DRAM if they wanted to. Why not do it?
Tesla's main goal is to do a lot while keeping the costs low. The fact the the entire interior is devoid of critical things like buttons is a testament to not how they're good at UX/UI, but how they've shaved off $$$. Pretty impressive. A potpourri of features in a Tesla car are all software-based (fixed NRE cost, essentially zero bloat on the COGS). Spending extra on a set of buttons, let alone a LIDAR system, would piss off the bean counters. If anything else, their marketing is great and people love the brand (and the stock).
Another argument is to not fuse the data. Just keep it as a backup decision maker akin to how Aviation systems are designed. If both systems agree, then you have additional reliability. Would love to hear from someone who's worked on these kinds of FSD systems.
Andrej covered this in a previous talk. If both systems agree, then nothing would be different if you did not have the 2nd system. If the systems do not agree then you ultimately still have to pick one in most scenarios.
IE if you are travelling down the highway at 80mph with a car close behind and your lidar+radar system says there's an immediate obstacle but your vision system says it's an overpass, do you emergency brake?
This goes for lidar and the interior, is simply what the believe the right solution is.
If they could save 100$ per car and have a traditional interior, I don't think they would do it.
The same for lidar, even if it was free, they wouldn't use it. They simply believe the additional complexity of sensor fusion is not worth it.
In regards to LIDAR - I think it's possible to drive safely without depth perception. One eyed humans do it just fine. It may not make sense to have LIDAR on every single machine you want to have vision processing with, including the Tesla bot. So it makes sense to solve problems without using it.
From the article:
> Develop the next generation of automation, including a general purpose, bi-pedal, humanoid robot capable of performing tasks that are unsafe, repetitive or boring. We’re seeking technical, electrical, controls and software engineers to help us leverage our AI expertise beyond our vehicle fleet.
Everyone knows robots are the future and it'll certainly be interesting watching tech companies entering this space. Apple is probably building one too (hiding it as autonomous car program).
I'm not really familiar with machine learning and how it differs from the type of work Tesla does with self driving software. I guess I was under the assumption some form of machine learning was being used for self driving?
Anyone have some insight on what he means by this? Maybe some examples of where we see ML being used today where it doesn't need to be?
Many repetitive tasks are better suited to a simple heuristic. In those scenarios using ML as a solution is like building a rocket to fly from SF to LA. In many cases, heuristics are both simpler and the best solution to a problem, but many companies are falling into the hype trap and using ML to solve problems where it's actually complete overkill or unnecessary.
I don't understand how that's "BS"? They're not marketing this to consumers.
It’s been said that AI is surprisingly hard because a lot that’s happening is unconscious, and once you fall into the rabbit hole of chasing it AI becomes a perpetual “5 years away” goal
I think that may be false and you might just not like Elon though.
A good example is Noam Chomsky and his political ramblings
Many people thought:
1. Elon Musk was a Silicon Valley software guy that couldn't possibly build electric cars or rockets.
2. The Wright brothers were a couple of bicycle mechanics that couldn't possibly build a real flying machine.
3. Steve Jobs was a right-place right-time washed up has-been that could not revive Apple and certainly could not make it the most valuable company on earth.
History is full of examples of smart and effective people transferring their abilities to new challenges.
As well as a good amount of non-bureaucratically/red-tape-restricted financial backing.
Those conditions seem highly likely to be able to produce results compared to other companies in the space - but I guess only time will tell.
For a funnier and maybe slightly less related example like yours though...
There is a surviving Japanese company that made samurai swords for a few centuries or so - they now make nail clippers, and they're some of the best damn nail clippers one can buy on this planet.
They then ventured into making tweezers as well. You would expect that if they can make such amazing and highly acclaimed nail clippers, they can probably make damn good tweezers as well.
Their tweezers are absolutely awful and they've never revised them for as long as they've been selling them.
As for “best and brightest” that won’t help as in terms of understanding cognition we are in pre-science stage, missing that qualitative jump from alchemy to chemistry so to speak.
all you need is a philosopher’ stone and a solid round of financing
Tesla shipped 500,000 cars last year, Ford shipped 2 million[1]. I know Tesla's valuation is sky-high, but I think that is the bullshit OP is referring to. To call them the singular successful US automobile manufacturer is ridiculous.
[1] https://s23.q4cdn.com/799033206/files/doc_news/2021/1/06/For...
https://pbs.twimg.com/media/E9FARebWUAYd5NI?format=jpg&name=...
Many of the production facilities that Ford uses, and huge sunk cost in things like engine production, are not an asset, but a cost in the long term.
Ford will not be making its own batteries, meaning somebody else will make a large part of the profit from each car.
Ford will not make its own self driving tech, so some other company will make that profit.
Ford will not sell its own cars, meaning they will lose large part of the margin on dealers.
Ford is massively invested in things that will be negative assets and has very, very little investment in the things that will matter in the next 5-10 years.
With 50% CAGR the PE ratio does not seem so crazy, if growth were to stop, then yes it is crazy.
PE is literally the worst way to evaluate a company that is growing.
But seriously, that is one of the single worst ways to evaluate a company and instantly discredits anybody that uses it as their primary argument.
I don't think it's out of the realm of possibility that, by the end of the decade, the Ford F-150 could become the best-selling electric vehicle. (Nor, for that matter, is that anywhere near certain. I withhold judgement on whether or not it is probable). If that were to be true, then Tesla's valuation really doesn't make any sense.
- Buy batteries from Tesla
- Buy FSD automation from Tesla
- License Electric drivetrain technology from Tesla
- Ford cars will charge at Tesla SuperChargers
Particularly with licensing it's possible that Tesla would make more profit on an electric F-150 than Ford!
Just announcing a electric version and producing a million of them are quite different.
Even if the F-150 electric becomes the best selling electric vehicle, that doesn't change my point either.
I didn't mean they're the singular successful US auto manufacturer
I meant that they're the first (for production scale that is, sure there are a few here and there making models w/ <1000 to be ever made) created in quite some time that has been successful.
I had no idea they had grown so big compared to other car companies.
https://www.theverge.com/2021/8/20/22633958/tesla-bot-elon-m...