Elon: Why Tesla FSD is MUCH HARDER Than LLMs
HTML-код
- Опубликовано: 6 окт 2024
- Head to geolog.ie/DRKN... or scan QR Code on the screen and use code DRKNOWITALL to get an exclusive 70% off of Geologie’s award-winning skincare trial set.
**If you are looking to purchase a new Tesla Car, Solar roof, Solar tiles or PowerWall, just click this link to get up to $500 off! www.tesla.com/.... Thank you!
Join this channel to get access to perks:
/ @drknowitallknows
**To become part of our Patreon team, help support the channel, and get awesome perks, check out our Patreon site here: / drknowitallknows . Thanks for your support!
Get The Elon Musk Mission (I've got two chapters in it) here:
Paperback: amzn.to/3TQXV9g
Kindle: amzn.to/3U7f7Hr!
**Want some awesome Dr. Know-it-all merch, including the YEAR OF EMBODIED AI Shirt? Check out our awesome Merch store: drknowitall.it...
For a limited time, use the code "Knows2021" to get 20% off your entire order!
**Check out Artimatic: www.artimatic.io
**You can help support this channel with one click! We have an Amazon Affiliate link in several countries. If you click the link for your country, anything you buy from Amazon in the next several hours gives us a small commission, and costs you nothing. Thank you!
USA: amzn.to/39n5mPH
Germany: amzn.to/2XbdxJi
United Kingdom: amzn.to/3hGlzTR
France: amzn.to/2KRAwXh
Spain: amzn.to/3hJYYFV
**What do we use to shoot our videos?
-Sony alpha a7 III: amzn.to/3czV2XJ
--and lens: amzn.to/3aujOqE
-Feelworld portable field monitor: amzn.to/38yf2ah
-Neewer compact desk tripod: amzn.to/3l8yrUk
-Glidegear teleprompter: amzn.to/3rJeFkP
-Neewer dimmable LED lights: amzn.to/3qAg3oF
-Rode Wireless Go II Lavalier microphones: amzn.to/3eC9jUZ
-Rode NT USB+ Studio Microphone: amzn.to/3U65Q3w
-Focusrite Scarlette 2i2 audio interface: amzn.to/3l8vqDu
-Studio soundproofing tiles: amzn.to/3rFUtQU
-Sony MDR-7506 Professional Headphones: amzn.to/2OoDdBd
-Apple M1 Max Studio: amzn.to/3GfxPYY
-Apple M1 MacBook Pro: amzn.to/3wPYV1D
-Docking Station for MacBook: amzn.to/3yIhc1S
-Philips Brilliance 4K Docking Monitor: amzn.to/3xwSKAb
-Sabrent 8TB SSD drive: amzn.to/3rhSxQM
-DJI Mavic Mini Drone: amzn.to/2OnHCEw
-GoPro Hero 9 Black action camera: amzn.to/3vgVMrH
-GoPro Max 360 camera: amzn.to/3nORGYk
-Tesla phone mount: amzn.to/3U92fl9
-Suction car mount for camera: amzn.to/3tcUfRK
-Extender Rod for car mount camera: amzn.to/3wHQXsw
**Here are a few products we've found really fun and/or useful:
-NeoCharge Dryer/EV charger splitter: amzn.to/39UcKWx
-Lift pucks for your Tesla: amzn.to/3vJF3iB
-Emergency tire fill and repair kit: amzn.to/3vMkL8d
-CO2 Monitor: amzn.to/3PsQRh2
-Camping mattress for your Tesla model S/3/X/Y: amzn.to/3m7ffef
**Music by Zenlee. Check out his amazing music on instagram -@zenlee_music
or RUclips - / @zenlee_music
Tesla Stock: TSLA
**EVANNEX
Check out the Evannex web site: evannex.com/
If you use my discount code, KnowsEVs, you get $10 off any order over $100!
**For business inquiries, please email me here: DrKnowItAllKnows@gmail.com
Twitter: / drknowitall16
Also on Twitter: @Tesla_UnPR: / tesla_un
Instagram: @drknowitallknows
**Want some outdoorsy videos? Check out Whole Nuts and Donuts: / @wholenutsanddonuts5741
All In Podcast with Elon Musk: • Elon Musk | All-In Sum...
Elon's post: x.com/elonmusk...
Head to geolog.ie/DRKNOWITALL or scan QR Code on the screen and use code DRKNOWITALL to get an exclusive 70% off of Geologie’s award-winning skincare trial set.
What is amazing for me is the way FSD ignores shadows and sun glare.
I hope they fix that shadow and glare problem with summons.
@@Saxafruge Was not aware that there was a problem
IMO, this is one of your best videos.
Great content!! Keep'em coming and do not be discouraged !! Thanks Dr. KIA!! Greetings from the sandy shores of Lake Huron!! Ontario , Canada !! GIDDYUP!!
I never got the sense that the idea was to force an LLM onto a Tesla. The idea will be for Tesla to incorporate the 'driver agent' of an LLM, which would be much smaller in scale.
You explained things very well and clarified the distinctions between the various technical areas very well. Thank you.
FSD is challenging? What? I thought FSD was solved. So what was Musk selling when he was selling $15,0000 vaporware? I smell a class action lawsuit. Didn't Tesla already go across the USA in 2018? I thought Musk said that was happening too? I get so confused.
Keep hearing how fsd is getting so close… yesterday it tried to turn into an oncoming car in an unprotected left. It signals way too early. It randomly pounds on the brakes or says it has taken corrective action without doing anything. It doesn’t remember what the speed limit was next day. Sometimes signals right when nav correctly says go left. Sometimes goes to get into left lane when in a block we have to be in right lane. In my area it still has a long way to go.
Same experience, and I do have some good days too. 12.5.2 stresses my wife out to the point that I only plan to use it alone for now. The future gets brighter.
March of nines
Tesla navigation might be better if it took direction from Google Maps. It always knows what lane to be in.
@@mnhsty I have seen many errors in google maps too. With Google maps when you report errors they fix them very quickly. Have twice now for me anyway.
Remember FSD only started rolling out full stack nueral nets 6 months ago. We are at GPT 1 right now with FSD and if you fast forward a few years it is going to be orders of magnitude better than it is now
FSD has been really good for me.
Completes 70% of my drives autonomously. The only interventions are mostly silly stuff like missing a speed Braker or not getting into exist lane which will be fixed pretty quick in the next few months.
Thanks John this session was very informative
Thanks for explaining these topics so clearly!
I would have loved having a professor like you when I was younger! Thank you for all the insight.
One token fp16 is 32-64 KB or even more if is not compressed....
To me it's the "filtering out" process that will consistently prove to be the most difficult task. How can the training model know up front what is relevant and what is not? Is it able to discern and identify that stop sign or stoplight from a cluttered background? Is that a pothole, and is it a hazard or not? This rabbit hole keeps getting deeper and deeper. How deep will it go? I was more optimistic about FSD being solved within two years, six months ago than I am now. Maybe all of the compute in the world won't solve FSD if the underlying assumptions and technical approach proves incorrect.
Great content!
if you want to run LLMs locally and efficiently, you need to develop new inferencing hardware. The current state of the art inferencing HW is not really usable for edge inferencing LLMs.
I disagree about the relative difficulty of LLMs vs FSD. One important point here is that, while language is already condensed down, that means an LLM actually needs to do a lot of background work to figure out what's meant by the words. An analogy that pops to mind is that FSD is compressing a file, while an LLM is decompressing a file. Both take a lot of compute! For example, an LLM just figuring out where the emphasis in a sentence is is super hard. Take the sentence "I never said she stole my wallet." and read it out loud multiple times, but putting emphasis on a different word each time. It completely changes the meaning of the sentence, right? You need to look back through everything that's been said previously for clues as to what is really meant here (since an LLM isn't listening to you, it can't hear where the emphasis was). This is just one of many many examples that make language quite hard to decode. Sarcasm, double entendres, innuendo, etc etc etc.
Were you able to get your wallet back?
@@DanHalper I never said she stole my wallet. Weren't you reading? 😉
The difference here is the criticality in the event of failure. For FSD, lives might be lost… for LLM, the worst thing that gets bruised is ego
@@milevets that's a different difference. The video was talking about the relative difficulty.
Funny how a couple of years ago Tesla channels were all touting how Dojo would kick butt and Tesla would be selling time on them.
Ho Hum.... meanwhile companies like NVIDIA are upping their game and can make there own at a cost that Tesla can't. These other companies have the facilities and foundaries in place to make these in quantity.
Dojo may be great for Tesla to use internally, but unless they sell the design, or have someone with the facilities to make them in quantity doing it, Dojo may only ever be an internal project.
2027 will be here in no time
alot of fabs popping up in the states
Dojo has always been discussed by Tesla as a high risk high reward project. Also, just because something is late doesn't mean it won't happen
@@5893MrWilson I'm sure they are working on it. However I think they will have mostly solved FSD using the NVIDIA stuff by then... along with their own HW4 inference chips. So Dojo may be more helpful with Optimus.
Has anyone looked at the AI4 board to see if it is designed to support a daughterboard? Seems like a good way to get a laptop sized model into the car that then refers to the cloud as necessary.
I would like to see this gigantic model running directly on the vehicle. Just curious in how the FSD would perform.
Context size: For me, "context" - statistical and reality variance, cross-sectional or longitudinal - is the absolute "key" to human or AI "competence." We learn to master the skills involved in a particular task through practice - making and correcting errors - while each "trial" occurs in an instance of a highly variable but "typical" context. To gain your private pilot's license, you practice flying and landing in variable winds, in rain or snow or sunshine or night, etc. Eventually your chances of failure or even rough landings diminish to the almost certain point of safety (given all the context clues) or the point of recognizing when not to take a chance. That we accept as a "level," indeed, a professional level of competence - at least worthy of a private pilot's license. And that really is a ticket to keep on learning. Humans do that (particularly while sleeping, converting short memories to long term). Tesla cars learn slightly differently, by gathering data (a huge bunch of it) to retrain the model and update the inference engine in the cars.
Gotta ask, out of 100,000,000 cells, how many are performing to spec, and what was the actual yield?
Absolutely. And humanoid robots are orders of magnitude harder than FSD.
Hopefully whenever they have to resort to cloud based services the functionality is adequate when no cell or wifi
What I can’t understand is why I don’t have smart summon?
You did a great job of ex-paining this. I missed the context point
It’s so much more difficult, but musk said it would be done 5 years ago.
Written language is even more compacted then spoken language. It loses all kind of information when compressed into setting.
👍
The Tesla inferencing chip is not better on a perf per watt basis than what Ambarella is selling
Why don't all these giant companies collaborate to put their compute together and use AI to cure cancer or aging?
Because it can’t model something never done before and then experiment to see if if the proposed idea works. It can be a tool in generating possible molecules, but they still have to be created abd tested.😊
Because it's gonna take decades and is not profitable. Recursive is having problems out of the gate using AI for this kind of work
Tesla's computational power significantly exceeds that of all other AI companies combined. Nonetheless, AI is still in its early stages and has a long way to go. This is primarily due to the human brain operation in quantum realm and at the same time minimal energy consumption make it the most efficient computation system exist in the universe. If a non-invasive method of implanting Neuralink devices were to become feasible, even in its current nascent state, AI could accelerate AI development and also serve human in its missions. I believe Tesla should collaborate with Starlink to access its supercomputer centers. Immediate calculations could be handled by inference computers locally, but more complex computations could be offloaded to the supercomputer which requires minimum latency. A partnership between Tesla, Neuralink, XAI and Starlink could achieve remarkable results.
King Pine
when you put a ad less than 5 mins into a 18 min video it encourages me to start training an ai plugin that can detect mid-roll creator ads and auto skip them (and im normally happy to see creators doing direct ads vs just youtube garbage, but this is feeling like yt garbage). heck, maybe someone already did.
8:24 and visual images. Most people look at a pen 🖊️🖋️ and know what it is.
Why is FSD so hard? Because it probably requires AGI, and AGI in the strict sense is quite possibly not achievable, though closer and closer simulations of AGI may be possible. AGI essentially means consciousness, and scientists have no idea how consciousness arises from material factors, or even IF it arises from material factors. There is no consensus on the issue. That's why it's called the "hard problem," though in reality it is the impossible problem, because material factors and consciousness are radically different from each other, with the result that scientists and philosophers cannot even imagine or conceive an observation that would make visible the bridge between matter and consciousness. To make this clear: suppose you are looking directly at a person's brain -- you are an observer at some sort of brain surgery. Now while you look at the brain, try to imagine how your eyes, aided or not by instruments, could actually SEE matter turning into consciousness. You can't, because consciousness is just not physically observable, light rays do not bounce off consciousness and impact the eyes or physical instruments. At most, one can observe things happening in the brain that correlate TO SOME EXTENT with conscious states. But that is not the same thing as actually observing the emergence of consciousness from matter, the transition between brain and consciousness. There is, on the contrary, a break in the continuity of observation in this case. To observe the relation between matter and consciousness, one must develop new, non-physical organs of awareness and then, rather than consciousness appearing derivative of matter, matter appears as a kind of coagulation of and expression of the non-material.
Doesn't this difficulty Tesla is having with compressing and feeding the tokens into the driving LLM call into question this AI approach? Originally, Tesla had the FSD software working just like a human driver. Classify what it sees through the car cameras. Then make a driving decision. The problem was that decision making approach required large amounts of C++ code. But now that AI makes it so much easier to write code, wouldn't that make it possible to deliver that large volume of decision making code?
What's being ignored is xAI is not just about LLMs, it's an AI company and Tesla FSD is also an AI project. I think he is under-representing what he believes xAI can offer Tesla in order to reduce the impression of a conflict of interest which obviously it is.
Starlink chip on an xAI cellphone on every Tesla, to facilitate communication with grok!!! xAI stays private.
HW3 it's at the end of the line, its not going to progress much more
What is the point of FSD? Your car drives you to work? Big deal. So does a bus, tram, train, subway and a taxi.
Ah, but you might have to share some of those with poor people. Never mind.
Next time could u talk about about Elon Musk an his12 years old attitude on X and in public
I read FSD doesn’t read road signs. That’s the primary reason for disengaging FSD. It needs to slow down to the speed limits on the signs rather than what the maps tell them. It is totally unaware of the road hazards that’s also on the signs. So often the Tesla drives too fast or doesn’t react to the driver in front of them.
In 12.5 it definitely sees road signs. I don't know if it can actually read but it definitely recognizes stop signs that aren't on the map such as stop signs held by construction workers.
Well all I know is, it’s called supervised for a reason.🤪
I have to disagree, Tesla has not got 12.5 to work correctly on Hardware 3 cars. My proof is v12.5.1.5 only went out to 5% of eligible FSD Hardware 3 cars and all push stopped dead after just 3 days. No new Hardware 3 cars have got FSD 12.5. While all eligible AI 4 cars have received 12.5.
I am willing to bet if and when I get FSD 12.5 a simple comparison with an AI 4 cars will prove that Hardware 3 cars continue to be erratic compared to AI4 cars.
I also can assure you those robo Taxis that will be shown off on 10-10 will be running AI 5.
FSD is 10 years old in 2 weeks - it started Oct 2014
and it's just SAE Level-2 certified - 10 years to barely get halfway - the easiest half of the task
it needs to be Level-4
and Robotaxi needs Level-5
It ain't going to happen folks - elon is just rearranging deck furniture on a sinking ship
What killed FSD?
1) Poor design
2) Poor equipment
3) Poor software approach
4) Poor management
Time to start over again elon.
_it's just SAE Level-2 certified - 10 years to barely get halfway - the easiest half of the task_
Your comment suggests you think that the SAE Level has something to do with how good, or how safe, or how skilled, or how useful a driving automation system is. But SAE J3016 makes it crystal clear that the SAE Level of an ADAS or ADS has nothing to do with its performance, and states explicitly that "[these levels] do not specify or imply that, for example, Level 3 is 'better' than Level 2". The *only* significant difference between SAE requirements for Levels 2 through 4 is who takes over if the ADAS or ADS encounters a condition it can't handle, and how long they have to take over. In Level 2, the human must take over immediately when requested. In Level 3, a delay is allowed, typically 10 seconds; and in Level 4 there's no requirement for a human to take over. That's why FSD is Level 2, while the vastly inferior MB Drive Pilot can be Level 3-it allows ten seconds for the driver to take back control when necessary.
@@BigBen621 FSD is Level-2 SAE certified and not L3 or L4 or L5 - that's what AI says.
10 years and FSD still does stupid things.
and of course my 2022 Model Y has hardware level 3 - ultra cheap cameras that can't read posted signs and will never make it to L4.
You can spew all the confusion you want but without LIDAR and cameras that can read at 20/40 levels FSD is a dead duck...
@@noleftturns _FSD is Level-2 SAE certified and not L3 or L4 or L5_
I agree that FSD is SAE Level-2. The problem is that you don't understand what this means. It means *only* that the driver must take control *immediately* when requested, in contrast with Level 3 with allows a significant delay or Level 4 which does not require a driver. You could verify this for yourself, but won't, because it would interfere with your false narrative.
_10 years and FSD still does stupid things_
And so does Waymo, and every other ADAS and ADS. And so does every human driver, many of whom have been driving for several decades. The difference, though, is that human drivers do stupid things far, far more often than FSD, as evidenced by the astonishing safety record of FSD recently released by NHTSA, compared with human drivers.
_You can spew all the confusion you want_
What I stated was facts. You could verify these for yourself, but of course that wouldn't fit your preconceived narrative.
_without LIDAR and cameras that can read at 20/40 levels FSD is a dead duck._
That is pure speculation on your part. Time will tell.
@@BigBen621 Mercedes has an SAE Certified Level-3 system
elon does not
it's that simple
Ask AI (Perplexity):
Q: does Mercedes has an SAE Certified Level-3 system?
AI:
Yes, Mercedes-Benz has an SAE Level 3 certified autonomous driving system called DRIVE PILOT. Here are the key details:
Mercedes-Benz DRIVE PILOT is the world's first and currently only SAE Level 3 system with internationally valid type approval.
It received certification for use in Nevada in January 2023, making Mercedes-Benz the first automotive company to bring SAE Level 3 conditionally automated driving to the U.S. market.
In June 2023, DRIVE PILOT also received certification for use in California, making it the first SAE Level 3 system approved for use in standard-production vehicles in that state.
The system will be available as an option on 2024 model year Mercedes-Benz S-Class and EQS Sedan models, with the first cars delivered to customers in late 2023.
DRIVE PILOT allows the driver to hand over the dynamic driving task to the vehicle under certain conditions, such as on suitable freeway sections with high traffic density, at speeds up to 40 mph.
When active, DRIVE PILOT takes over tasks like controlling speed, distance, and guiding the vehicle within its lane. It can also react to unexpected traffic situations independently.
Mercedes-Benz takes on the liability when the system is active because the vehicle has assumed what they call "dynamic control"
.
It's important to note that while this is a significant advancement in autonomous driving technology, the system still has limitations and requires the driver to be ready to take control when prompted by the vehicle.
=================
elon is just plain lazy - or too busy stumping for Trump.
@noleftturns
I don't and never have disputed that Mercedes has Drive Pilot, which is certified SAE Level 3 in Nevada and California. Here are the limitations of Drive Pilot:
● It operates only on geofenced, precision mapped, limited access highways
● It operates only at speeds of 40 MPH and below
● It operates only in the same lane-no lane changes
● It operates only with a lead vehicle within 100 meters ahead
● It operates only in clear weather-no rain or snow
● It operates only during daytime hours
● It operates only with no direct sun on the cameras
● It operates only if the driver doesn’t avert his gaze from ahead for more than five seconds
In essence, Drive Pilot operates only in traffic jams on freeways, during the day, in dry weather. Meanwhile, Tesla FSD operates on every street, road, highway or freeway in the U.S. and Canada, mapped or unmapped, at any speed up to 85 MPH, with or without any other traffic on the road, with automatic lane changes, day or night, in almost any weather, in any sun conditions.
As I've pointed out before, the main difference between Level 2 and Level 3 is that in Level 2, the driver must be able to take over immediately when requested; while in Level 3, the takeover can be delayed-for instance Drive Pilot allows 10 seconds for the driver to take over. This is trivially easy when all vehicles are moving in the same direction at
Tesla pumper channel
rocket man is a traitor