Musk: Tesla No Longer COMPUTE CONSTRAINED!!
ฝัง
- เผยแพร่เมื่อ 24 มี.ค. 2024
- Elon Musk, Ashok Elluswamy and others just gave us more clues that 2024 will indeed be the year of Embodied AI. From Tesla finally passing beyond compute constraint hell to FSD 12 becoming semantically aware, things are looking great for both Tesla vehicles and their Optimus Teslabot!
**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/referral/john11286. 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 AI STUDENT DRIVER Bumper Sticker? Check out our awesome Merch store: drknowitall.itemorder.com/sale
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 TH-cam - / @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
Source Posts:
/ 1771649240302850158
elonmusk/status/1...
elonmusk/status/1... - วิทยาศาสตร์และเทคโนโลยี
![[UNCUT]”หมอปลา และทนายแก้ว” ลากไส้ขบวนการ เด็กเชื่อมจิต จ่อเอาผิดทั้งครอบครัว I คนดังนั่งเคลียร์](http://i.ytimg.com/vi/9zXrAIVBQFo/mqdefault.jpg)
Condolences on your father's passing. I only catch your channel once in a while and sadden to see anyone losing a parent from their life.
My model 3 with 12.3 was impressive. But while driving through a home construction area it did not know what to do with the huge dumpsters sitting in the road. It stopped and just sat there. I hate dumpsters taking up half the road, hard for me to see around them. I had to take over until I was out of where the dumpsters were. Otherwise it did an amazing job, for the first time my Tesla took me out of my church parking lot and drove me home. It was not able to handle an intersection I too hate till yesterday.
Ashok may be a Nobel Prize candidate if FSD gets any better. It could become a REQUIRED safety feature in ALL vehicles. It should.
We need this. After decades of downward pressure on deaths per VMT, smartphones reversed that trend. Humans can no longer be trusted behind the wheel. I think it's fair to give the human the option, but most will prefer to be on TikTok than drive.
What about Elon?
Thank you for pointing this possibility out- I never connected the dots myself that FSD could likely become a safety requirement!
Elon won’t be considered for any prize. Bad Elon.
Sorry about your dad, glad you could make it for the memorial. I lost my mom a year ago and am so happy I made it
even though far away.
Linguistic point: as I understand it, grammar is about sentence structure, semantics about meaning.
Additional earnings reports from major tech companies, driven by Nvidia, coupled with trader FOMO, could fuel a resurgence in market buying pressure. I'm considering investing over $300k, but I'm uncertain about risk mitigation strategies.
Nvidia stock is roaring like many did during the 1990s bubble. But this time around, the hype around new chips is happening in a more mature demand environment.
Certain Ai companies are rumoured to be overvalued and might cause a market correction, i think it’s best you reach out to a proper fiduciary for guidance
De-risk your portfolios, shore up your core holdings, and take some profits while balancing your portfolio allocations. I’d also suggest you go with a managed portfolio, but even those don’t perform so well, so it’s best you reach out to a proper fiduciary to guide you, that’s what works for my spouse and I. We've made over 80% capital growth minus dividends.
this is all new to me, where do I find a fiduciary, can you recommend any?
She goes by ‘Heather Ann Christensen’ I suggest you look her up. To be honest, I almost didn't buy the idea of letting someone handle growing my finance, but so glad I did.
As a data scientist i cannot emphasize enough how important not being computer constrained is.
Once you have your data quality, sampling algorithms, and everything else in place, then the sky is the limit as long as you have enough compute.
A lot of companies make the mistake of hiring a small team of expensive data scientist and then they just don't allocate any budgets for compute. Having constant access to large amount of compute, not just for that one big model training, but constantly being able to run all the experiments you want to both at large scale and smaller scales is super valuable. Being able to run many experiments concurrently helps you iterate so much faster.
In addition to the jump in smoothness, I have also observed a couple of big changes in v12.3. Previously in v11.x, I used to set navigation to a couple of normal destinations such as grocery store. 11.x would set a course that would take me past the destination and then requiring a U-turn so I would override that by dropping a pin on the map and making a more logical left turn. With v12.3, nav still sets the less efficient route but when I get there, the car departs from the mapped route and takes the turn that I previously had to designate with the dropped pin. It doesn’t re-map, it just ignores the mapped route on its own. I’ve had it do this on two separate destinations so far.
Does anyone know why sat nav algorithms do this sort of thing? It’s extremely annoying especially when driving a large vehicle that’s difficult to reposition.
Greetings from Athens too, the initial one though P:
Sorry for your loss.
I just got V12 yesterday, and I have to say For the first time in a long time, It feels like they'll get there. And soon. V12 is two quanta above V11, easily. Everything works better, and everything seems to work. I have disengaged a few times, because I didn't like the lane it was in, but not once yet, has it done anything remotely "wrong". On the last version, I'd have to disengage 3-6 times per drive.
Absolutely amazing!! Thank you Dr!!
So sorry for your loss.
To me the hardware limitation is that there is nothing to keep the cameras clean. FSD works in snow but when it gets really bad some of the cameras get occluded (same with mud). The rear camera becomes useless pretty quickly in the winter, so backing up on FSD in the future will be problematic. It would be a similar issue if front cameras were added.
Your absolutly right, and this proves a limitation problem by human thinking even if they are "Elon" smart. Living in Calif. or Texas. MUD, SNOW, WINTER Etc.
My condolences 😔
Thanks for the lovely video and I agree with you. 😊
What happens if there is a power outage, and there are no traffic lights. Just traffic?
Maybe the driver turns off FSD?
Elon is the man!!
I dont think Elon was talking about HW3/4 when he said hardware has it's limits.
I think he was referring to Lidar and Sensors!
Software/AI is the answer.
I do love your content but videos gave been getting so long. Delighted to see a video not too long to watch. Thanks 😊
The short term solution is that, with Grok. You can verbaly say... "approaching Construction zone" then the car can approach with caution ad have a diff algo.
"or have the laws change" where theres got to be ,"Construction Zone" signage on ALL const. Zones, the car will read it and approach meaningfully.
Interesting. Thanks.
Chuck Cook has his left turns. Digital-Dan has his overpass/underpass intersection. So far, v12 has mastered one of the two difficult "right turns" in this intersection. The second I still had to disengage, but honestly I don't even know what the right way to handle that one is.
❤ condolences
Wow! WOW! Wow!
I agree!
My condolences on your father's passing. Prayers for you and all your family.
Great video
Anyone else with FSD 12.3 have the issue where the car just doesn’t go fast enough? I have to constantly press the pedal to force it to go the speed limit.
12.3 is great though!
Yes! It'll sit at 30 in a 40 zone, and just won't accelerate!
how will the fsd handle flashing red lights ?
Thank you for this explanation. Is Robotaxi Level 5 ?
A Full on *Human Taxi Driver, isn't even exactly "Level 5!" They get Distracted, Make Judgment Errors, Fail to Observe Realities that can affect the Decision Making, and get in Accidents!
Is "Level 5" Autonomous expected to have "Fewer" Accidents, or "Zero" Accidents? Can it be expected to "Avoid an Intersection Related T-Bone Crash?" Etc!
My condolences on the passing of your father. May your family and loved ones find peace during this time!
Condolences!! I had a question. There is a guy that posted some videos of FSD12.3 that failed to react to an aluminum ladder on the road. He tries different positions but rarely FSD does the correct response. Any thoughts why this may be the case?
Excellent FSD progress coverage. I would venture that this advancement will lead to Optimus becoming the first "human 2.0" to go to Mars. It just makes more sense.
You are not real right?
Of course robots were in the Mars plan all along, to prepare the physical requirements of the first settlers. It's not like the Western expansion where water, game and the elements for construction were sitting there when settlers arrived. That all must be manufactured or synthesized from raw materials before humans set down.
keep on pumping captain dunning krueger
Sorry to hear about your father's passing. My deepest condolences to you and your family.
How often will software updates be released then if they're no longer compute constrained?
Assuming you answer the privacy settings to share data. Are you able to determine if any video clips from your car are being uploaded to Tesla?
Interesting
Aw. Sorry about you losing your father. My condolences to you and your family.
Great info. I’m exited. I’m also scared. More excited, but also very scared.
lol humanity will be compute constrained ad infinitum
Amazing. I can't wait to buy one when there is a tv and lounge seats instead of a steering wheel and pedals.
Well no construction zons allowed if FSD cant manuver it.
The computer is finally using Socratic logic. That is a HUGE change to how humans learn.
Elon said sometime ago already, that the FSD computer is aware of the world around it and learning more each day from the flurry of data from all the Teslas driving around.
Saw a Tesla add on TH-cam this afternoon.
L5 cannot happen with existing fixed point vision sensors. Already got into a $400+ mishap, not sure, if it was due to compromises made in accuracy and latency in the FSD V12.3, to make it efficient using the limited FSD compute on the vehicle. Dojo, will soon be able to justify where the hardware is limited, if not help overcome it.❤👍
Elon Musk doesn't need to advertise, but if he did! The mind boggles.
The good thing about these Nvidia GPU chips is the latest is double the 100 and it is plug and play. So adding compute capacity is very cost efficient.
My condolences to you for your father's death. I must have been mistaken as I thought it was your mother who was close to death. I certainly hope it was not both. Your Dad must have been very proud of you and what you have achieved. Keep up the good work. Best wishes from the UK.
The bots could learn to drive ice vehicles and load and unload them....sweet.
Can I take your explanation of the iterative learning method of FSD and BOT Intelligence as an introduction to AGI? I expect I am not intelligent enough nor knowledgable enough to phrase that question correctly. So what I am getting to, is would you consider doing a long form master class on the subject, like through several intensive videos? With respect to this one video, I think I understand AI better than from all the other videos and reading material I've previously tried to absorb.
🙏
The big question for me is will the existing camera hardware on the cars actually support self driving to the required level? I think this is all about camera placement and reliability. Given the fact that apart from the internal camera the external camera placement and number hasn't changed since the first model 3 came out Tesla must be very sure that they have enough. I have my doubts - similar to Chuck with the side cameras being too far back. I also note that in Europe traffic lights are usually much closer and higher up for the cars at the front of an intersection - my car often just doesn't see them.
If the cameras aren't right no amount of software intelligence will be enough and Tesla will probably be either retrofitting or refunding.
There’s no competition for fsd but there now seems to be a lot of competition for the bot so it’s a good thing it doesn’t have to split resources anymore
These one-off feats of magic don’t offset the flurry of flubs on easy maneuvers we see. At least not for me.
Lately i got interested in crypto but has no idea on how to go about it. How does it work please..
As a beginner, it's essential for you to have a mentor to keep you accountable. I'm guided by Fergus waylen a widely known crypto consultant :
Riding the market wave entails understanding the market volatility, Fergus Waylen has been doing an awesome job reviewing the charts, enabling me to capitalize on the volatility via day trading.
I am surprised that this name is being mentioned here, I stumbled upon some of his clients testimonies on CNBC news last week..
What impresses me most about Fergus Waylen is how well he explains basic concept of winning before actually letting you use his trade signals. This goes a long way to ensure winning trades.
His technical analysis is excellent and hid interpretation/projections of the market is so accurate I sometimes ask myself if he is human haha. Point is, Waylen is the perfect trader to follow for advise and daily signals.
Tesla all the way :-)
Does this mean their Dojo supercomputer project was a failure?
When do you think the car will talk to us?
It's SHOCKING how many times I've posted the correct pronunciation of Ashok's name on this channel.
Hint: It rhymes with choke.
How do I know? I've listened to his own self introduction.
Come on, don't leave us hanging, what is the SHOCKING number of times?
I believe Elon may be alluding to some sort of secured commitments for the new NVIDIA GB200 Blackwells.
They're a ton more powerful than the old H100 stacks, and would definitely remove compute constraints for any company that can acquire them.
I'm not convinced that HW3 and HW4 have limits that mean anything at this point. Chuck is worried about the .01% situations, which arguably is about 1,100,000 miles per day based on 11 billion miles driven in the USA per day if all vehicles were autonomous. I honestly think we are in such early days for the End to End story that saying the current hardware isn't enough is a foolish mistake.
I look at nature and even the smallest of life can operate intelligently enough for path planning with very little power. So, I think the inference demands will diminish overtime not increase. It may increase before decreasing, but in the end, I think the path planning will become a much simpler NN than today.
What is the resolution of a human eye?
A lot more than Elon's $10 webcams
Compute is a verb, dammit! It all started with third quarter spend. Ugh!
Chuck Cook discusses FSD hardware limitations in a video titled "2 Problems FSD V12 Cannot Escape" on the Hans Nelson channel (around 1:15). He believes that hardware 3 (and presumably 4 as well) can't reach level 4. He discusses the position of side view cameras behind the driver's head (especially when the driver leans forward) and the lack of depth perception for example.
Tesla's ahead on Tesla's approach (tautologically), but Waymo's ahead on its approach, and it actually has robotaxis on the road in limited areas, and it's not standing still while Tesla proceeds on its path. Google can certainly do massive computing and AI. Tesla has no lead in these areas generally.
This is a problem for the existing cars only, more cameras (just two or three more?) would solve that.
a few cameras, and several lidars, and lasers, along with distance sensor should solve it.@@bluetoad2668
@@bluetoad2668 More cameras could help but also require retraining, and existing cars provide the training data. Lidar and other sensors could also help.
@@restonthewind - Question: Which "Other Sensors?"
The Video mentioned "High-Def Radar", for example.
Other ideas?
For my thinking, placing the Cameras on the "A-Pillars" (Left & Right) could do a lot for Object Distance Calculation, especially due to the "Wider Base Distance" possibly allowing more refined "Angular Acuity" for the Triangulation Calculation! (Than the Typical few inches a Human's Eyes are separated!)
@@robertweekley5926Waymo uses lidar, radar and optical sensors. Tesla used radar and ultrasonic sensors in addition to optical sensors before 2022/23 models.
Rationing compute must bias alot of the tech work towards efficiency and limit experimentation
Will be interesting to see how things go with a more back access to training cycles
Kinda like the old stories of dozens of staff sharing access to a universities ine punch card machine
I lived through punch cards as a student and partially as a programmer ten years later. We wouldn't have been compute constrained unless the DOD wasn't sucking up all the machine cycles every night when we were trying to test. Case Western U. in 1968
Any idea why Musk was so circumspect in recent months about dojo chips? The way he spoke it almost sounded like the chip was not designed appropriately for the current software training.
It’s not. Teslas had to purchase massive numbers of NVIDIA chips to do most of the neural net training they need. Dojo may still have a role but it’s not clear as yet
Hmm. I thought it was just that they couldn't get Dojo fast enough & Nvidia filled a big gap. Has anyone admitted another problem?
Could be I/o contrarians, could be memory constrained?
Please let Elon know that older FSD owners would like to get v12 before new owners. My condolences on your father's passing.
When will FSD become FSD?
Or said otherwise, when will it require No Supervision?
When?
I am super excited by the tech and the progress, but if the answer to that question is Never, then it will not help TSLA investors.
Optimus, on the other hand will…
The observed differences in both structural and functional neuroanatomy between ketamine users and controls may explain some of its long-term cognitive and psychiatric side effects, such as memory impairment and executive functioning. Given the effect that long-term ketamine exposure may yield, an effort should be made to curb its abuse.
All still good, through a difficult time🙏💔 Best in AI/Robotics/Manufacturing, ...you know what it's about🧐😂👍
This also has massive implications for Tesla One and the rate of innovation Tesla can sustain.
A question on LiDAR vs Tesla vision (and it is coming out of place of lack of technical knowledge on my part)
Why do people consider lidar and “intelligence” as mutually exclusive?
Tesla vision uses camera to scan the surrounding. And then uses machine learning to understand semantics of it.
Other companies that use lidar and sensors can also do that, right? Use lidar and other sensors to scan the surrounding. Then use machine learning to interpret that data.
But everywhere I see, they talk like companies using lidar cannot use machine learning, by saying things like - “it’s not just about scanning things. It’s about semantics”. I am like - “Yeah, they can scan and understand semantics the same way Tesla does. Only difference is the way things are being scanned. “
Suspicion: LIDAR "Resolution" is Blockier and Non Colored Image Resolution, providing less precise object identification, maybe!
Was/is "Color" important in some things? Well, obviously, Traffic Lights. Possibility that it plays a role in other things, too.
Sensor conditions: High Resolution Cameras vs LIDAR Resolution, how do they compare? And, how does the Data Processing System, decide between the Two Image Types, as to Accuracy and Relevancy?
Basically, is one more "Blind" to some things, than the other? And, if so, which "Driver" would you be more comfortable with "Driving? The one with Better Vision, or the one with Poorer Vision?
@@robertweekley5926 ahh ok. Thank you for your response.
what I want to know is if I can send the tesla to the curbside pickup at McDonalds, where it sits in a preassigned slot, with the worker walking to the parking lot and hands the food to the tesla, and the tesla drives the food back home to me. Can tesla do that? thx.
Of course it can't.
When your car is driven by a generally intelligent robot with materials handling who is also your financial agent, then it will. I'd say ten years, give or take. I'll take a bet I could afford to lose on that. 😎
Tesla will eventually reach Level 5 but it will take a long time to get enough data on those very rare edge cases.
I can see a new market emerging here, intelligent road pylons that not only have a visual signature but maybe have a a WIFI or bar code or such that can tell a self driving EV what it is there for.
For example in the maritime world buoys use the color green and red to denote which direction a channel is in relation to the buoy.
And not all pylons need to be smart, say one has a WIFI and the others just have color markings and such.
And the incentive for construction companies to employ these might either be due to laws but maybe insurance incentives as well in that if a poorly laid out set of pylons caused a self drive EV accident the construction company might be sued.
Also one might develop smart road signs as well, that have WIFI that can tell a self driving EV more details about a section of up coming road that otherwise would be too hard to show graphically on the sign.
For example there might be a problematic lane change that could be denoted by such a sign .
Also even cars themselves might have WIFI that can alert other cars, such as if a car breakdown in the middle of a road.
And these could be useful for any car be it self driven, and EV or ISE car or whatever.
The idea is that not all the burden need be put on the self driving EV's AI but the road infrastructure could be made more self driving friendly as well and that might be for non self driving EVs as well!
In regard to sensors there is one sensor that might need to be added , that being audio in that it is generally considered dangerous for a human to drive with headphones on so that they can't hear sound in the environment such as a siren, horn blast, or strange wounds from ones car or wheels or a loud thump if one hit something, so as not to make a hit and run and such.
But here we have things like audio cameras that just require an array of microphones and some processing.
e.g. Google
Acoustic cameras can SEE sound
And another approach for certain tricky sections of road is that instead of fixing the AI in general, those spots might be tagged in a database and treated as special cases by the AI with no need to create a special physical sign post for them or the like but instead create virtual ones.
A couple of more thoughts.
As we now might have Google Street view as Tesla's drive along the roads they might build a virtual twin of the highways they drive. These would not have the actual traffic on these roads but could have a statistical treatment of them vs time of day and such.
Then before a trip the Tesla could simulate the trip in a simulate mode where the car is in park and feeds the simulated data into the front end of the sensors as if it were real data.
BTW, that has been done in developing avionics systems where the avionics is tested on the lab using simulated sensor inputs and works really well!
Thus if one is going to make a trip that day, the night before one might have their car make a simulated trip as a preview of coming attractions. And this could be used as a norm to compare with the actual trip in that say the traffic is different than one expects that could be part of the logic in how to drive that section where it is different.
Also in this seem mode the driver might sit in the car and practice drive it over certain sections to show the AI how they the human would drive that section such as a tricky lane change or the like.
What about GERMANY, SELF DRIVING FROM TESTLA 😅
Levels are easy if you visualize as follows.
What is needed to drive
0 legs, hands, eyes, brain, body & physical vehicle controls needed
Not needed
1 legs
2 hands
3 brains (awareness)
4 body (driver)
5 Physical vehicles controls
5:42 instead of 1 forward camera, they should have two. In windshield corners. And perfectly parallel. It will allow software to create binocular vision - allows better depth perception. Knowing the exact distance between the cameras and the angle of an image in the view allows it to triangulate quickly
Don't you think they would have done this already if there was any benefit to it? Cameras were and are cheap components. Don't forget, computers don't get distracted and they can calculate distances many many times a second. And its accuracy does not have to be one millimeter. A normal human driver's accuracy is around 2 inches...
When approaching an unfamiliar object or one of an unexpected size, binocular vision is valuable beyond ordinary visual distance estimates. FSD already does a good job fitting through tight places, but bincular ranging would be useful at greater speeds and distances, and probably in reduced visibility situations.
@@kafiluz4317 it would require different algorithms and programming but ultimately would be more effective. It’s simply not possible for it to not be more effective.
They probably figure they can cut into the difference with good software. But regardless, it would be better. At a minimum it’ll give better field of view without having to distort the image
As long as you loved your parents and not have siblings that cared more about the $$$ than my Mom’s wellbeing. My family had one bad apple. To this day I failed to give my Mom the best ending. I did my best and my sister called the police on me 3 times.
Dojo chips was probably their worst investment. They will end up not really using it at all, too hard to use compared to nvidia ecosystem of software and hardware.
Every thing In our world in intelligence problem
Time to break out some dry powder and add a few more shares. Tesla is getting close, and moving faster as it starts to close in. :-)
Tesla may be ahead but an analogy I would use it Tesla started out a snail's pace, then as AI tool got better and the chips got better they could move at an ant's pace, then a rat's pace, then a human's pace, etc. Companies that are joining the race late are able to jump in at whatever speed technology is currently going at. They didn't need to go through the snail period for example. And then you ask how far Tesla got in the year they spent at the snail's pace and it's maybe a few steps at the human's pace. And a year at human pace won't get you that much farther ahead when we get to a car's pace. And a year at a car's pace won't mean much when we get to a jet's pace. Eventually we'll have so much compute and such good AI tools we'll be able to build a new AI from scratch and get it driving a car within a few days. So you can't say Tesla is "years ahead", you can only say they're as far ahead as it currently takes to build an AI and get it driving, and that time is shrinking. If the day we can build and train an AI in a few days is maybe 4 years away, then that's the most Tesla could be ahead. Next year they can only be, at most, three years ahead, then two years ahead the year after that, etc.
_So you can't say Tesla is "years ahead", you can only say they're as far ahead as it currently takes to build an AI and get it driving, and that time is shrinking._
You're missing the fact that Tesla has billions of miles of actual driving video, which no one else has and no one is likely to be able to replicate in any reasonable time.
@@Big_Ben_from_La_Mesa and you're missing the fact that many companies are working on AI models that don't need that. Did you need billions of miles of driving to learn how to drive? No, right? AI in the future we'll need far LESS training data than a human. A few examples and verbal explanations will be enough for it to understand what to do. That's exactly what I mean by getting better AI tools. The #1 thing all these robotics companies are working on is the ability to show it what you want it to do and it will "get it" with one or two demonstrations and be able to jump in and do it perfectly.
@@truhartwood3170 _Did you need billions of miles of driving to learn how to drive?_
That's a false equivalence, on two levels. First of all, humans have general intelligence that allows us to learn to drive without billions of miles of training; FSD and other ADSs don't have the machine equivalent, AGI; and AGI may never be achieved for ADSs, given the limits on mobile computing power.
Second, human drivers cause >44,000 fatalities a year just in the U.S. One goal of ADSs is, or at lease should be, to dramatically reduce this death toll; and simply replicating human driver training in ADSs won't achieve this.
The reason billions of miles of training is important is that the greater the mileage used in training the model, the more edge cases will be experienced; and the fewer accidents, injuries and fatalities will occur because the ADS encountered an edge case it hadn't been trained on, and didn't have AGI to figure it out on its own.
_The #1 thing all these robotics companies are working on is the ability to show it what you want it to do and it will "get it" with one or two demonstrations and be able to jump in and do it perfectly._
In most cases, they are dealing with known processes and situations; for example in manufacturing. In the ADS field, we will *never* be able to completely describe "what you want it to do", because that will change minute to minute on the 4 million miles of roads, again just in the U.S. We will only be able to approach this asymptotically, and only by driving billions of miles. And we won't know until we get there just how many billion miles this takes.
@@Big_Ben_from_La_Mesa Yeah, see, you're talking about the way things are *right now*. I'm talking about where things *will be*. The reason people cause accidents is usually not because they don't know what to do when they see a sofa on the road and so they just drive straight into it or steer into a truck or something. They cause accidents because they're distracted, looking the wrong way, too slow to react, are impatient or road raging, etc. Someone who drives for a living isn't significantly safer than someone who just commutes and drives a bit on the weekends. You're the one making a false equivalence. What I'm saying is that it obviously doesn't take billions of miles of driving to learn how to do so.
Also, I'm not saying they won't need a significant amount of driving data, but probably just thousands of miles instead of billions. And they won't need to drive that in real time, they can train on the data in days or hours. And the AI in the future will probably be way better at understanding what the point is, so they won't need as many edge cases. They see maybe a dozen videos of avoiding objects on the road and they understand the concept and can rapidly apply that to any object. Same with humans. Show it a few videos of someone hitting an object when avoiding it would hit a human and it understands that hitting objects is better than hitting humans, done. AI of the future will have enough reasoning and understanding to extrapolate the takeaway lesson and then generalize that to any situation it comes across.
People (like you) so often look at the way thigs are right now and think that the problems we're encountering today need to be solved with brute force application of the tools we have today. Like back in the 1800s or so when people thought computing devices and/or robots would be super complex gear mechanisms like in mechanical watches because that's all they knew. Would we ever be able to pack enough batteries and compute power for self-driving if transistors hadn't been invented and we were still using vacuum tubes? Of course not! For decades AI work was stalled with Convolutional Neural Nets and Recurrent Neural Nets (CNNs and RNNs). It went almost nowhere until transformers were invented in 2017 at which point it took off. Then came stable diffusion just a couple years ago that allowed another huge leap forward. My guess is we'll start to see the gap between major discoveries shrink now that so much research and money and focus is on it and there's so much competition. What will AI look like when we have a dozen more leaps as big or bigger than transformers or stable diffusion were? What about when we have toolboxes of open-source AI nets for vision, physics, motor control, language, etc? Expand your mind. People like you thought computers would just be better typewriters and the internet would just be a faster way of sending mail. Some aspects of AI in the near future will be unimaginable to most people today. And the world 30 years from now is something I don't think anyone alive can really picture (well, assuming humans are still thriving, people are pretty good at picturing all the bad outcomes).
@truhartwood3170 (misposted to you, sorry!)
❤💔💔❤❤❤ Sorry for your lost. It's so hard, I know.
And I hope you will be Ok after some time.
Come to france fsd is still a disappointment
Why are you assuming, with what appears to be near certainty, that Tesla has not solved the problem Yann spoke of?
Only Tesla knows for sure, and some v12 behavior makes me think they have solved it.
After years of railing on Tesla’s design choices for FSD they finally got it right. End to end NN is going to be the winner.
Ketamine is the only remaining constraint.
Elon is unconstrained by compute for perhaps the first time in his life. All things are possible now.
Unfortunately he's STILL constrained by our corrupt government.
Well if its not hardware constrained they must just be bad at writing software, because it clearly does not work and its making the same mistakes it did 2 years ago.
Personally I think its not an either or question. If you are more intelligent you can live with less information (less or lower quality sensors). Even when people are legally blind, the human brain learns to adopt and overcome.
But conversely if you have better sensors you can live with lower IQ. This is very relevant for the current state of AI. This is why I think Tesla took the wrong aproach. It will take too long for AI to reach this level.
👍🚀🛰👍😎
You gotta stop saying "While you're down there" at the end of your videos. It has other connotations. 😊
No longer compute constrained? That doesn’t matter it’s still bullshit. Tesla has not activated FSD 12.3.x on 2021 and older model S and X. Something it’s because it doesn’t have the internal cabin camera like Model 3 and Y… but in reality it’s showing that the new way that FSD v12.x operates is not as safe as they claim. If it’s really infinitely better than the previous FSD 11.X, then it should be an automatic upgrade for every car, not simply held back on Teslas most expensive car models S and X because FSD 12 needs more time!!!
Omg, like any claim of musk’s can be taken at face value. Everything he says is to pump up stock and hype up investment, and he’s been busted staging demos towards that end multiple times.
Nobody and nothing can save Tesla!Its game over for Tesla!
You sound vaccinated.
I just ordered my cybertruck Foundation model with all the bells and whistles for an incredible amount of money. Hopefully this will be coming soon to cybertruckland
FSD is very impressive, but my dog can translate Chinese better than that! ;-)