Ashok Elluswamy
@aelluswamy
Followers
49,111
Following
603
Media
15
Statuses
561
Explore trending content on Musk Viewer
Renzo
• 149514 Tweets
#WWERaw
• 82544 Tweets
#GodMorningTuesday
• 37096 Tweets
Scarlett Johansson
• 33379 Tweets
Oilers
• 32756 Tweets
Fani Willis
• 31365 Tweets
定額減税
• 30436 Tweets
Kingdom Hearts
• 28122 Tweets
राजीव गांधी
• 26553 Tweets
梅雨入り
• 25308 Tweets
Birds Nurturing
• 24042 Tweets
給与明細
• 24010 Tweets
Gunther
• 22163 Tweets
#Canucks
• 20882 Tweets
Vancouver
• 20788 Tweets
#ファンパレハーフアニバーサリー
• 18934 Tweets
Edmonton
• 17581 Tweets
ドジャース
• 17426 Tweets
金額明記
• 17231 Tweets
Amber Rose
• 15464 Tweets
ショートアニメ
• 14937 Tweets
キングダムハーツ
• 14089 Tweets
Lyra
• 12452 Tweets
国民実感
• 10659 Tweets
Presented some of the recent work from the Tesla Autopilot team at CVPR this year, especially about "Occupancy Networks" - our approach to solve general obstacle detection and using it to enable sophisticated collision avoidance. Full talk here: (1/12)
147
720
6K
High-fidelity park assist is shipping this weekend to Tesla customers without ultrasonic sensors as part of the holiday release!
304
429
5K
This end to end neural network approach will result in the safest, the most competent, the most comfortable, the most efficient, and overall, the best self-driving system ever produced. It’s going to be very hard to beat it with anything else!
257
592
4K
This is the sort of driving that's really hard to code explicitly, but our end-to-end approach brings in almost effortlessly.
FSD V12 avoids standing water
(I know hand held phone videos are a no no, and I’m sorry, but I had to share)
165
245
2K
254
392
4K
It’s been my secret ambition to make Tesla FSD so good, that it works as well as a local driver even in India. Which will then prove unequivocally that we have achieved AGI :)
Great meeting you today
@elonmusk
! We had multifaceted conversations on issues ranging from energy to spirituality.
5K
25K
207K
482
451
4K
Motivating to see the overwhelmingly positive response for our latest FSD v12 release! Kudos to the incredible team at Tesla, it's an honor to work with them everyday!
207
360
4K
Whether some surface is drivable or not is all contextual. We typically don’t want the car to drive on dirt, but since the paved road is closed, the car needs to drive on dirt. The same logic also applies to, for example, mounting small curbs to avoid a larger obstacle. So
128
423
4K
Handling many real-world situations, like construction, requires understanding and reasoning about the scene semantically and not just geometrically. That is why self-driving is an intelligence problem and not a sensor problem (i.e. needing LIDAR, Radar etc).
FSD V12 handles a chaotic construction section like an expert, achieving smoothness I couldn't have managed even on my first attempt.
96
200
1K
196
464
3K
Efforts over the past several months went into making this new approach beat the previous v11 stack (developed over many years). We believe v12 has net surpassed it now, and hence the next set of 12.x releases should bring unprecedented progress.
171
345
3K
We are basically building AGI, but with video and control as primary modalities, instead of text!
Tesla AI is building next-generation autonomy on a single foundation video network that directly drives the car
Join the team and build state-of-the-art end-to-end models using massive fleet data on one of the world's largest training clusters
449
1K
6K
343
297
3K
It’s a no-brainer for safe and convenient day-to-day driving, and it only gets better from here.
You can now subscribe to FSD (Supervised) for $99/month in the US
Upgrades > Software Upgrades > Subscribe
2K
2K
14K
181
290
3K
Neural networks + cameras work amazingly well and are really the solution to robotics. Same solution for the car, for Optimus and will be for all such artificial animals.
Optimus can now sort objects autonomously 🤖
Its neural network is trained fully end-to-end: video in, controls out.
Come join to help develop Optimus (& improve its yoga routine 🧘)
→
3K
8K
35K
91
251
3K
Many collisions prevented along the way
135
179
3K
Pulling these off requires extremely accurate prediction in space and time!
There is zero room for error here. 😬
The ability for FSD v12.3 to judge if an opposing car is going to stop or go on a yellow light is exceptional. This left on yellow is one of the most risky maneuvers to execute and I’m so glad that the team has worked on getting this
95
276
2K
141
278
2K
And it took a tremendous amount of high quality data from the fleet, a very large amount of compute and a world class engineering team to get to this point!
106
117
2K
Nice tests!
BREAKING: does FSD 12.3.3 react to hand waves? You be the judge 😏
Huge shout out to
@bilalsattar
for willing to stand in the line of fire today 😬🫡
@elonmusk
@tesla_ai
@tezmeeinc
161
357
3K
89
198
2K
My daughter loves it too, keeps making me do Autopark over and over.
New autopark feature is crazy. We had way too much fun with this.
59
83
1K
152
137
2K
This is the first version of FSD that uses neural networks "end-to-end", i.e. going from raw camera videos and vehicle kinematics to producing control actions to drive the vehicle.
33
122
2K
These robots do use additional sensing such as audio, proprioception, temperature etc., just like most animals. But no LIDAR, RADAR, ultrasonics, HD maps shenanigans. It will become so obvious over the coming years, we’ll be wondering why anyone thought those were a great idea.
93
209
2K
Tesla has just the perfect setup to solve robotics, with expertise across the full stack starting from actuators, batteries, manufacturing, super advanced AI software, efficient inference hw and soon 100 exaflop of training compute! It's the best place for any engineer to be rn!
Tesla is building the foundation models for autonomous robots
1K
4K
22K
51
218
2K
Surreal watching these grow better!!
There’s a new bot in town 🤖
Check this out (until the very end)!
3K
7K
32K
158
112
1K
These predictions are already used to prevent a lot of collisions. For e.g., Autopilot prevents ~40 crashes / day where human drivers mistakenly press the accelerator at 100% instead of the brakes. In the video Autopilot automatically brakes, saving this person's legs (7/12)
44
270
2K
If saving lives, while building super cool things excite you, please consider joining us:
60
192
2K
This technology progression is very exciting for us, because, when we execute on this plan we can build a car that literally never crashes. It will be an absurd thing for a car to crash. (10/12)
35
163
2K
This is the v1 release of this technology, and will have follow up releases that have even better geometric consistency with the cameras, better persistence of occluded obstacles, etc. For now, enjoy parking 🤭and happy holidays!!
86
78
2K
Many of these improvements go to Tesla drivers in the 10.69 release later today. All credit goes to the amazing team that works hard to make this happen. (11/12)
40
89
2K
Updated thread with a lot of cool maneuvers!
Reply with your best FSD video that you’d show to someone who hasn’t experienced V12 yet 👀
2K
2K
9K
58
126
1K
The Tesla Autopilot team is at CVPR this year! If you are in New Orleans, do stop by the booth, talk to the engineers & checkout the Cyber.
42
126
1K
Wow, incredible!!
26
70
1K
Another example of this
47
78
795
23
66
1K
Hate is for the weak!
Arnold
@Schwarzenegger
has a powerful message for those who have gone down a path of hate.
211
2K
5K
146
128
602
Occupancy Networks ♥️ NeRFs
The occupancy representation of these networks allow for differentiable rendering of images (based on the Neural Radiance Fields work). However, unlike typical NeRFs, which are per scene, these occupancy nets generalize across scenes. (6/12)
22
95
937
"Occupancy Networks" - our solution to these problems - predict volumetric occupancy of all the things around the car. i.e. every voxel or continuous point in 3D space has a probability of being occupied and also its future motion. (3/12)
10
60
813
@RohitP_tweets
It’ll be faster over the next few releases. Was tuned conservatively for the first release.
44
59
787
This allows the car to avoid any UFOs on the road. Also the car can peek around the corners to make human-like unprotected turns. (5/12)
14
45
700
This replaces the 2D obstacle band that customers had with a high-resolution 3D reconstruction of the Tesla's surroundings. This is an extension of our Occupancy Network, with much higher resolution to help with tight parking maneuvers.
8
24
662
In addition to obstacles, we also predict painted lines on the ground, also in 3D. Together, these help perform the full parking maneuver just by looking at this one screen.
12
28
659
Think ~2 µs for a query compared to a few 10s of ms for a classical baseline. (9/11)
10
26
591
These Occupancy Networks have some nice properties:
- Predict geometric occupancy
- Use multi-camera & video context
- Predict dynamic occupancy flow
- Persistent through occlusions
- Resolution where it matters
- Efficient memory and compute
- Runs in ~10 ms
(4/12)
4
27
578
The obstacles are modeled as a continuous distance field. This allows us to represent arbitrary shapes in a smooth and computationally efficient way. The vehicles you see are not some fixed meshes, but the network's real-time prediction of the shape.
5
17
571
ha ha
14
27
546
Lot more collisions can be saved but requires more smarts to be both smooth & fast while being safe. Implicit neural networks help with this planning problem by providing quick queries for probability of collision. (8/12)
4
24
519
We have a relatively small, but highly talented team of engineers. The ratio of engineer to compute is already one of the best. The generative model featured here was just two engineers hacking for a couple of weeks.
These imagined futures can be action-conditioned to produce different outcomes.
For example, the videos below are generated entirely by the neural network by simply using different prompts
32
240
3K
29
37
517
It was great hosting Lukas last fall. I'd say that this is the typical experience at the Tesla Autopilot/AI team. If you are interested in building cutting-edge real world robots consider joining us:
10
57
511
Typical approaches such as image-space segmentation of free space or pixel-wise depth have many issues. (2/12)
2
19
507
For all investors in earth
22
53
436
Will be on shortly!
28
23
436
Excited to see this approach replace the more traditional object / instance detection approach. Helps predict samples from high dimensional distributions.
📢Introducing Pix2Seq-D, a generalist framework casting panoptic segmentation as a discrete data generation task conditioned on pixels. Works for both images and videos, with minimal task engineering.
work w/ Lala Li,
@srbhsxn
@geoffreyhinton
@fleet_dj
3
53
344
13
28
279
lol
9
17
249
God, the White Lotus show is so haunting! The characters and the soundtracks keep playing in my head…
29
24
186
Was looking forward to something like this!
15
24
189
Some engineers from the team are at Vancouver right now for CVPR '23. Stop by the booth to know more about the team or if want to get a demo drive.
6
12
152
This!
Continually torn between wanting my kids to have opportunities I never did and not wanting them to be spoiled brats
115
103
3K
4
2
106
@chapman_bliss
@neuralink
That would be amazing. Your findings can probably help accelerate AI progress a lot. Especially interested in understanding how the brain simulates actions and predicts consequences efficiently :)
4
1
92
Was playing homophones with my 4 yr old. I did ear-hear, wood-hood etc., then she goes: daddy, chicken’s a food and chicken’s a bird! 😬 🤭
5
2
80
Wow! Wicked fast NeRFs
Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
paper:
project page:
github:
21
359
2K
2
6
61
Good start to the day, lol!
Wordle 212 3/6
⬛🟨⬛🟨⬛
🟨⬛⬛⬛🟩
🟩🟩🟩🟩🟩
9
0
51
Surreal what’s happening in Afghanistan. Can’t believe that there’s such brutality and raw violence-based control of power in this day and age!
4
1
51
@Naveenkwin1
Unfortunately, we are a relatively small team and fully in-person. That said, exceptions can be made for truly exceptional people.
0
1
17
@bratwurstkomet
Any kind of prediction of future outcomes depends on generative models (since it's almost impossible to explicitly annotate every outcome, you only have a sample at a time)
0
0
5
http://www.youtube.com/watch?v=mVqlrU33md4
@krishashok
suppress and oppress and depress a software developer..
0
0
4
Excellent essay. Nails the issue : Nuclear power is our gateway to a prosperous future:
http://t.co/vUFgSmzy
0
0
3
Check out the cool college guys of the 1913's in Chennai...
http://www.flickr.com/photos/79806301@N00/4773208755/
0
1
4
Just spent 45 mins helping a relative install Skype over phone. And the mission is successful !!
1
0
2
You know you hate comic sans if you hit delete before even reading mail. Especially the maroon coloured ones
1
0
1
Just registered for CEGAM 2011 - CEG Global Alumni Meet July 03, 2011 at
http://cegam2011.doattend.com
1
0
2
Che che che... Bakery ah develop pannalum pannen.. Bannu venum venna venum nu uyirra edukuraanga appa...
#vadivelJokes
0
0
2
Hilarious RT
@cricinfo_news
: India races to get Nehra, Munaf unfit in time for semi-final | Page 2 | ESPN
http://es.pn/hmNhgE
0
0
0
No the name "Indi Commandos" was not bad enough for the kochi team.. Now they are called "Kochi Tuskers Kerala"..
#worstname
0
1
0
Prevention against Radioactive meltdown.. Pls RT:Iodine pills, Japan and science.
http://bit.ly/fNPona
0
0
0
@shrutihaasan
You must then see Godzilla in Tamil
"Odunga! athu nammala nookki thaan varuthu" :P
0
0
1