비전 인코더는 시각 정보를 숫자 형태로 바꿔주는 역할
FSD와 옵티머스 시각 정보를 인코딩 잘 해야하므로 둘다 사용할수 있게 고도화 개발한듯
“vision encoder shared between FSD, Robotaxi, and Optimus…trained on millions of internet videos”
There are going to be humanoid robots EVERYWHERE in big cities the next five years and it is going to be completely insane and world changing.
1
반대로 말하면 다른업체들도 어느정도 수준까지 데이터를 모은다면 천장이 있겠고 일반성능은 어느정도까지 올라올수 있겠네요
문제는 아직도 제대로 데이터 수집하는 업체가 없다는 것이지만
Replying to @Scobleizer
Tesla has Niagara Falls amounts of data. So much, that we will keep much less than a millionth of it. Unless the data is of a super weird, long tail event, it’s irrelevant.
7
"테슬라는 나이아가라 폭포에 비견될 만큼 엄청난 양의 데이터를 보유하고 있음.
데이터가 너무 많아서 그중 100만분의 1도 안 되는 양만 보존할 것임.
데이터가 극도로 특이한 롱테일(long tail) 사건에 관한 것이 아니라면, 사실상 이제 의미가 없음."
Replying to @Scobleizer
Tesla has Niagara Falls amounts of data. So much, that we will keep much less than a millionth of it. Unless the data is of a super weird, long tail event, it’s irrelevant.
1B, 3B 작은 모델들로 이정도가 되는구나
Replying to @BLKMDL3 @wholemars
In case anyone is curious:
We’ve had a ~pixel-perfect world model since last year, which is critical for synthetic data generation.
The (much) smaller version of that in the car is a conditional mean-flow model, based on the recent sequence of Tesla Universal Vision Encoder features, which are the same ones that are plugged into the 3B encoder model that produces the driving actions.
We fork off from these UVE features, then run a separate ~1B diffusion model. This takes in the previous reality-view generation with added noise and then denoises it to the current visualization.
It then produces a small video chunk (8 images at 36 fps, at the resolution of the UI screen). This work is done by 1/2 of a TRIP engine every UVE frame at 18 hz and is spread over 4 cycles.
Currently, the latency is a bit high due to compute constraints, but we have ways to interleave newer UVE features along the diffusion transformer layers to cut roughly 35% of the latency.
1
17
Replying to @BLKMDL3 @wholemars
In case anyone is curious:
We’ve had a ~pixel-perfect world model since last year, which is critical for synthetic data generation.
The (much) smaller version of that in the car is a conditional mean-flow model, based on the recent sequence of Tesla Universal Vision Encoder features, which are the same ones that are plugged into the 3B encoder model that produces the driving actions.
We fork off from these UVE features, then run a separate ~1B diffusion model. This takes in the previous reality-view generation with added noise and then denoises it to the current visualization.
It then produces a small video chunk (8 images at 36 fps, at the resolution of the UI screen). This work is done by 1/2 of a TRIP engine every UVE frame at 18 hz and is spread over 4 cycles.
Currently, the latency is a bit high due to compute constraints, but we have ways to interleave newer UVE features along the diffusion transformer layers to cut roughly 35% of the latency.
Theorem retweeted
테슬라 FSD v15 포트홀 회피, 기대 이상
- 생각보다 훨씬 똑똑하게 피함
- 다른 포트홀 구간 찾아 추가 테스트 예정
cf. v15 릴리즈 노트에 포트홀, 도로 세부 감지 개선 포함
Theorem retweeted
(일렉트렉) 테슬라 FSD v15, 가장 큰 약점들을 정조준
- 평가 : 리얼리티 뷰는 진짜 인상적, 주행 개선이 화면만큼 진짜라면 큰 업데이트
- 개선 항목 : 표지판, 포트홀, 스쿨존 속도 등 지금 FSD가 가장 약한 부분을 거의 다 겨냥
- 25배 : 핵심 개입률 개선, 진전을 재는 올바른 지표지만 실제 수치 없이 배수만 공개
- 계산 : v14가 핵심 개입 사이 약 2,000마일이었다면 v15는 약 50,000마일, 제한적 무감독 서비스 기준(약 10,000마일)을 훨씬 넘음
- 우려 : 고속 주행 프로필에서 운전자 모니터링 완화, 메이저 버전마다 두 걸음 전진 한 걸음 후퇴 반복
생각보다 호의적이네요.
electrek.co/2026/11/21/tesla…
Theorem retweeted
어제는 무슨 말인지 몰랐는데, 오늘 확실히 알게됨
Replying to @roger_kappler
Yeah.
The new user interface will re-render reality in a mind-bending way. Hard to describe, but it’s awesome.
Theorem retweeted
테슬라 FSD v15, 개입 시 인간 피드백 삭제 이유 (추정)
- TRV : 3D로 재구성된 장면으로 개입 이유를 AI가 직접 파악 가능
- 개입 횟수 : 위급 개입 25배 감소로 일일이 물어볼 필요 줄어듦
- 피드백 품질 : 귀찮아서 아무거나 누르는 경우가 많아 오히려 잡음 (사용자 불만도 많았음)
Theorem retweeted
FSD v15 영상에서 가린 화면 : 향후 릴리즈 기능
- 질문 : 왜 전비(26.4kWh, 100mi) 화면을 가렸나
- 테슬라 AI팀 테리 왕 : 그게 전비가 아니라, 향후 릴리즈에 나올 다른 멋진 기능들임
cf. 가린 부분이 무엇인지는 미공개
Theorem retweeted
가우시안 스플랫과 디지털 트윈 : 신뢰를 높이는 시각화
- 핵심 : 차 카메라 실시간 영상과 미리 만든 도로 디지털 트윈을 신경망이 합쳐 도시를 실시간 렌더링
- 확장 : 다른 테슬라, 옵티머스, X 라이브 영상까지 합쳐 화재 현장, 맛집 대기줄 같은 변화를 실시간 공유
- 로보택시 : 앱에서 목적지 줄 길이, 입구 위치까지 보여줄 것, 1년 안에 가능하다는 예측
- 옵티머스 : 집 디지털 트윈에서 화분 물주기, 오븐, 차고문 점검 같은 작업 지정
- 경쟁력 : 플릿, Grok, X, AI 팀을 모두 가진 곳은 테슬라뿐
결국 신뢰가 게임이고, 옵티머스로의 확장은 엄청 설득력 있음
cf. 일론머스크 : 꽤 정확함, 테슬라 차량이 교통 정보를 공유하고 Grok의 실시간 X 피드와 결합해 최적의 이동 경험 제공, 목적지에서 무슨 일이 일어나는지도 미리 알려줄 것
cf. 가우시안 스플랫은 2025년 10월 ICCV에서 애쇽이 처음 공개, 카메라만으로 약 220ms에 3D 장면 재구성
Gaussian Splats lead to more trust.
A neural network makes the digital twin around you that Tesla just announced/demonstrated here.
Or, rather, joins the real world that it is collecting from your car's cameras and mixing it into a digital twin it has already imaged of the road ahead and real world around you.
This is a software platform that will let Tesla continue to run ahead of anyone attempting to compete for years.
People keep asking me why Robotaxi will be better than Waymo, or Zoox, or other vehicles that come out of China.
Because soon the neural network will gather road data from other vehicles and join that into the visualization too.
Imagine a house fire starts in San Francisco.
A Tesla rolls by. Images the house fire for its own display. But it knows that is different than the "ground truth" digital twin of the house that it captured sometime in the past.
It shares just that house and fire with the network.
All of a sudden, if you are somewhere else in the city and you see a plume of smoke you can zoom in on the display to see what's causing that. And there is a visualization of the house burning.
Then the firetrucks arrive.
And another Tesla rolls by. And captures the changes happening, and shares that with the network.
Then the first responders start laying hose, and closing streets, around the fire, and start working to extinguish the fire.
Everyone can watch to see what's happening in that neighborhood as more Tesla's roll by.
And someday Optimus will join the digital twin. So, if an Optimus walks by the scene, that data gets put onto the network.
When my neighbor's house caught fire I watched all this. And the owner was out front recording his house burning before the trucks even arrived on his phone.
Imagine that someone capturing it for X was live streaming it here.
That visual data could be incorporated too.
Everything you see on the visualization is a voxel.
Computer vision can watch all voxels and understand the changes in the real world.
And report it to everyone.
And can even incorporate first reponder radio from somafm.com/scanner/
So everyone knows what is going on and can see it on their phones, glasses, or Tesla screens.
What was just announced is a new way to render the city in real time.
Imagine you can zoom into your favorite BBQ or Burritto joint and see how long the line is before you even take a Robotaxi there.
In Austin I took a Cybercab to one of its BBQ places and when I got there learned there was a two hour line.
Imagine that the Robotaxi app would be able to show you, in real time, the line and estimate it accurately based on historical data.
Back in 2005 a Microsoft Researcher showed me a system that could predict how long it would take you to get to Seattle over the bridge based on historical data. That saved a lot of people pain going across the bridge, but it wasn't real time.
Between Microsoft's campus today and Seattle are hundreds of Teslas on the road right now.
This is why I say the fleet is what matters from here on and why others won't be able to catch up.
No other robot platform has Grok. Has X. Has an AI team that is executing on this this way.
My consumer research shows that trust is a huge reason why people will switch from Uber to Tesla, or Waymo to Tesla.
The one that shows consumers more things about the city around them will gain more trust.
When I took a Cybercab to Austin City Limits a few weeks ago it didn't tell me how to get to the entrance of the festival. Or what it looked like. Or how long the lines were. Or how to get a pedicab to travel the mile from where it dropped me off to the front gate.
I predict that it will within a year.
And I only trust my life to Tesla's AI every day. Waymo every once in a while. (If you use an autonomous vehicle you are trusting your life to AI).
That trust will grow as it shows us the real world in more and more detail.
And only Elon has all the pieces to increase that trust.
Now, when Optimus comes? It will do the same in your home.
Imagine you have 15 plants, like we do in our home, to water.
Some need watering every day. Others need watering every week.
Imagine walking around a digital twin of your home and clicking on each plant and telling it "water this one every week."
Among many other things, like doing a security detail. "Check the oven after we leave and make sure it is properly turned off." Or "when we leave, check the garage door, gate, and front door are properly secured."
All of this work leads to this.
It is why I don't see any real competition coming for Tesla robots.
Trust is the game. Trust improves with better safety and visualizations.
All this will be obvious to everyone within two years. Especially in cities of the future that allow autonomous vehicles.
And being in such a city will even improve your life. You'll be able to pull out your phone and see what's going on in your city even if you don't have a robot or use an autonomous vehicle.
The data will come from a Tesla rolling by. Or a Tesla owner standing in line. Or an Optimus robot walking around doing tasks for its owner.
The fleet is the game. No one else will be able to match it in America.
내비에 익숙한 사람들이야 기존 찰흙 뷰로도 대충 찍겠지만 그렇지 않은 많은 일반 사람들은 이게 정말 유용할듯 함.
로보택시를 개나소나 아무나 쉽게 탈수 있게 대중화 시키겠다는 의지로도 읽혀짐
Replying to @jamesdouma
It is mesmerizing to watch.
Where this is actually useful is when you want to tap the exact drop-off or parking location on the screen that you want your car to target.
44













