Raines @raines1220 Twitter profile

Pinned Tweet

Raines

2 days

Sentiment Analysis of Actually Smart Summon (ASS) Key Statistics * Total results analyzed: 100 * Positive results: 72 (72.0%) * Neutral results: 17 (17.0%) * Negative results: 11 (11.0%)

2

0

26

Last Seen Profiles

@___stannnn

@MoKusanagi

@katherineap15

@Flamion_

@littlebigvoid

@Un762023New

@stw_pdg

@SophiaHunger

@wselir

@NathsnR

@CCunningham76

@cyborked

@KarenCupps3

@sungjinfancams

@GreenvlHabbo

@JamesBrock1350

@Fashtao

@l3linha

@Daily_Nina_Anna

@Finance_Bareek

@naomajimetw

@CALA3So

@Pasutri69639203

@emmitt8x

@SukaIbuIbuTua2

@Flamion_

@TyseleyEnergy

@DarrenModzelew1

@BabylonLiving

@ccarolcarreira

@favmorenita6

@RobMFriend

@AnaMonsalveA

@tomatodamenako

@openwindow_offl

@The_Word_

Raines

@raines1220

19 days

Elon arguably fulfilled all his promises. 1. v12.5.1.3 is ~14x better overall and ~4x safer than v12.3.6 2. This month, FSD 12.5.X already got ~1.4x better overall and ~2x safer

43

89

881

Raines

@raines1220

21 days

I was exhausted after the gym and didn’t notice the cat at all. If it weren’t for FSD v12.5.1.3, I might have had my first roadkill this year. But FSD saved the day, and now I feel really proud.

10

31

304

Raines

@raines1220

4 months

@TheSonOfWalkley DON'T tell those bears, but I have some secrets for bulls. "亲爱的特斯拉车主感谢您对特斯拉的长期支持!我们很高兴地通知您，我们即将推出一项重要的测试活动，旨在为早期购买全自动驾驶软件(FSD)的车主提供特别体验。您受邀参加我们的 FSD

33

34

250

Raines

@raines1220

4 months

I've tried FSD v12.3.6 for three days. 99.9% of my drive is handled by FSD. When I usually walk 100 meters to the mail room to pick up my package, I use FSD. When I want to change to a different parking spot, I use FSD to drive out and then use autopark to find a new spot. I try

16

26

244

Raines

@raines1220

4 months

@ArtemR I doubt that this is actually FSD driving because, currently, it seems that FSD would show a red wheel error and stop itself in this heavy fog. However, let's assume it is FSD driving and, for some reason, it feels confident. Then, I'm interested in discussing this. First of

32

5

216

Raines

@raines1220

3 months

I randomly came across a video that got me curious about how Figure01 was implemented. When it tried to pick up an orange, it missed but kept following its original plan. After finishing, it seemed to realize the task wasn't completed and tried again. It looked like rule-based

33

24

185

Raines

@raines1220

3 months

基于我的估值模型，特斯拉的转折点在于何时发布2万美金左右的车型——无论是CyberCab还是Model 2。我大概估算了一下，北美网约车每英里的成本大约1.5美金，比私家车大约0.5美金贵3倍。网约车业务每年行驶约400亿英里，仅占全美3万亿英里的1%左右。根据我的估算，Robota

36

20

171

Raines

@raines1220

4 months

@Tesla_Cybercat 有次在你的Space中我好像提过：算力解决后其实更新速度没有加快太多。小版本十天，中版本二十天，但之前也是如此。我当时分析，大概率是因为瓶颈在于验证，而验证流程不可避免，没想到被Elon确认了。

23

31

162

Raines

@raines1220

2 months

This post confirms many of our previous guesses: 1. The reason for ‘lane dancing’ in v12.4 is that normal driving got ‘filtered out,’ leaving only object avoidance clips, making the model believe more lane changing is better. 2. The ultimate issue is an intelligence issue,

Elon Musk

@elonmusk

2 months

@SawyerMerritt Sorry for the delay. This release had far fewer interventions, but suffered in driving smoothness. Part of the issue was too much training on interventions and not enough on normal driving. It’s like a doctor training too much on patients in the emergency room vs training on

402

405

4K

17

3

154

Raines

@raines1220

19 days

FSD v12.5.1.3在统计数据上已经达到人类驾驶近似的安全水平。假设今天特斯拉就用FSD v12.5.1.3来发布无人出租车，我相信它每百万公里事故数量和人类相比会处于同一个数量级。我相信我的这个观点会有很多争议，但是我对这个结论有很高的信心。欢迎一切辩论。

28

6

117

Raines

@raines1220

2 months

They said Baidu is way ahead of Tesla for Robotaxi. Well, I agree because, for now, FSD cannot beat human driving. If you notice from 8 to 13 seconds, the remote operator's hand is not on the wheel. This indicates that the remote operators are supervising the car even when it is

20

19

114

Raines

@raines1220

2 months

FSD v12.5 is not 100% “photon in, control out” for three major reasons: 1. They need to implement some emergency heuristics between the neural network and actual control to deal with worst-case scenarios. v12.5 still occasionally fully stops at blinking lights even though the

15

5

101

Raines

@raines1220

4 months

@onenewbite 我做过深度研究，把大多数b站上测试华为ADS 城区NCA的视频都看了一遍，把无加速、无剪辑、露大屏的视频基本上逐帧分析了一遍，目前的结论是：FSD

16

11

93

Raines

@raines1220

19 days

@TSLAFanMtl The community tracker. Unlike others, I got my data from the 'tester view' and compared how the experience changes within the same groups of testers (so it doesn't matter whether they prefer to disengage more or less).

7

1

91

Raines

@raines1220

4 months

The information on v12.4, v12.5, and v12.6 shared by Elon is probably much more valuable than most people think. It fundamentally changed my overall understanding of FSD and my confidence in achieving full autonomy. I've lost a huge number of brain cells since I thought deeply

20

10

89

Raines

@raines1220

23 days

I estimate that FSD (unsupervised) will be approved in Q2 of 2025. Real-time tracker:

12

10

84

Raines

@raines1220

2 months

WTF, MY FSD JUST REVERSED!!! Here's a dash cam footage where FSD saw a yellow light and slammed on the brakes, which is bad. Since there was not enough braking distance, it stopped on the crosswalk. Then something creepy happened. FSD v12.4.3 reversed a little bit to stay away

10

2

80

Raines

@raines1220

25 days

[1/2] What an "amazing" drive I just had—TWO critical disengagements on my first drive with v12.5.1.3 (~5 miles). I'll do my best to describe what happened, along with my best attempt at root cause analysis. In the first one, the car braked very late in front of a moving train,

10

2

77

Raines

@raines1220

19 days

I just realized, from a statistical point of view, that FSD v12.5.X is already on the same order of magnitude in safety as humans. This is insane—how stupid I was not to realize this!

8

1

76

Raines

@raines1220

2 months

I'm an optimistic guy. Based on currently known information, HW4 with a 4x better camera and a 3.5x better inference chip, bundled with a 5x larger next-gen FSD model, means that a profitable Robotaxi business is already achieved, assuming a team of remote operators at a

6

5

74

Raines

@raines1220

1 month

I am a nobody with a 2024 Model 3 Highland with 'advanced software update' turned on, and I got 12.5. Note that currently, according to Teslafi, about one-third of the 12.5 cars are not Model Ys.

9

2

75

Raines

@raines1220

2 months

I always keep in mind that the FSD team is undoubtedly the best in the world. In this regard, there's still a bright side to seeing them struggle and delay due to some issues in v12.4. I know with certainty that other competitors (e.g., Huawei, XPeng, Li Auto, etc.) are willing

Chuck Cook

@chazman

2 months

It's Monday and SFSD v12.4.2 never arrived. Sure enough the @Tesla ADAS Operators are out validating a new version. More than likely we are up to v12.4.3. I have seen three different vehicles so far today, no Cybertrucks. In the images you can see the approx 10 minute loop.

47

25

609

8

11

73

Raines

@raines1220

4 months

After several days of heavily using FSD, I've become such a bad driver. Sometimes I don't know who has the right of way or which lane to take. I'll just give up: "FSD, now you should disengage me and take the wheel. Otherwise, I'm going to crash." My wife just got her permit.

6

5

69

Raines

@raines1220

25 days

People now use FSD much more than they did with previous versions. It seems their demand was "activated" by the better FSD.

6

4

68

Raines

@raines1220

2 months

I cleared my schedule for tonight, and I'll read every character and frame from everyone's post if it relates to FSD 12.5.

8

0

69

Raines

@raines1220

4 months

I analyzed all causes of FSD disengagements and grouped them into two categories: easy and hard. Easy problems are reflexive, requiring less intelligence, such as stopping at stop signs. Hard problems involve more intelligence, such as reasoning, prediction, or subtle

3

7

67

Raines

@raines1220

4 months

@elonmusk I have a gut feeling that v12.3.X is intentionally not prioritizing the reduction of critical disengagements, thereby making most of us analyze something that is irrelevant for now. Maybe, all the disengagement improvements we’ve observed are just “side effects,” just like the

5

3

63

Raines

@raines1220

19 days

This is why I estimated that FSD is already on the same order of magnitude in terms of safety as humans. The current problem is that, unlike humans, FSD's mistakes are not evenly distributed across all situations and locations. Once the worst cases of FSD also reach human-level

13

7

65

Raines

@raines1220

2 months

Based on my experience with FSD v12.4.3, I already envision a future where FSD is fully solved on HW4 cars with a model that is 5x larger. On AI5 cars, with a model that is 50x larger, it drives 10x safer than humans. I estimate that Tesla will officially solve FSD on whichever

Ray

@ray4tesla

2 months

Why is TSLA up so much right now on Robinhood? Did I miss anything?

210

45

1K

6

4

65

Raines

@raines1220

1 month

In a previous study, I explained that FSD intervention data can show statistically significant differences due to varying disengagement tendencies among different testers. To accurately reflect real-world changes in FSD across versions, the most crucial adjustment is to compare

13

11

64

Raines

@raines1220

2 months

TBH, I think FSD won't go anywhere without HW4. Balancing safety and comfort does not do anything meaningful. Maximizing intelligence should be the first principle. Do we still remember the giant leap in v12.3? I believe that is the point where we maxed out most of the inference

27

3

62

Raines

@raines1220

4 months

@smdcapital88 Lol, I forgot to post. After seeing Elon’s post, I got some insights that eliminated some significant risks that restrained me from buying more $TSLA. But it’s not super cool analysis, just my random thoughts. Originally, I tried to make it sound really cool just for ��clickbait,”

4

10

60

Raines

@raines1220

2 months

After gaining much more experience with v12.4.3, based on my personal experience, the overall need for disengagement/intervention has decreased, ranging from approximately -20% to 50%. This version involves changes in almost every scenario. Some scenarios improved, while others

7

3

61

Raines

@raines1220

4 months

@PGStheone Well, I generally think the Chinese market is more open-minded with fewer anti-tech people there. The Huawei ADAS system has a take rate of over 80%, selling at ~1000 USD, even though it only achieves <7 miles per disengagement in city roads (based on my own research). The

3

1

61

Raines

@raines1220

1 month

We might have solved lane issues. In v12.3.6, 32.83% of disengagements were due to lane issues. Now, we’ve collected 11 disengagements from v12.5, none of which are lane issues. This means it is ~98% likely that we have improved lane issues. In fact, based on current data, we

7

4

60

Raines

@raines1220

4 months

@onenewbite 我随便发两个我做过的工作，分享一下为什么我会如此“口出狂言”： (1) 【一刀不剪云观光问界M9挑战北京城区智能驾驶-哔哩哔哩】 —————————————— 7:18 司机表示避让逻辑比较合理，没有傻等的情况 7:36-8:03 被白车、行人、公交车吓住开始傻等，司机偷偷干预油门共计6次

10

6

56

Raines

@raines1220

15 days

According to the NHTSA and VTTI, 80% of collisions and 65% of near collisions involve some form of driver distraction within 3 seconds of the event. According to the Zendrive collision report of 86,000 collisions, 57% involve phone use, and nearly 17% of all crashes involve

5

57

Raines

@raines1220

1 month

友商从来只在FSD上面做加法，从来不做减法，这就说明了问题。友商添加激光雷达，却不敢去掉摄像头。友商添加VLM，却不敢去掉端到端驾驶模型。从逻辑上来说，这证明了FSD方案的正确性，却证明不了他们“改良”方案的必要性和有效性。因为有效的方案里都有FSD的骨架，而“改良”的方案中改良点却各有不同。这

TSLA99T

@Tsla99T

1 month

我才搞明白，理想的这波主动接受采访，原来是为了和HW ADS

10

0

52

11

4

56

Raines

@raines1220

2 months

Tesla is creating a world where I wake up every morning and realize that life is getting a little bit better. I also find myself staying up less because I can’t wait to meet tomorrow.

6

2

57

Raines

@raines1220

2 months

如果8/8延后，对股价的影响其实是很恶劣的。听我和你细嗦。首先，每多延后两个月，FSD的表现可能就会变化数倍，如同v12.2-v12.3的数倍，也如同v12.3-v12.4的数倍。这会导致远程接管的数量出现严重负增长，极其恶劣的导致运营成本出现负增长，从而消灭很多潜在的工作机会。

13

1

54

Raines

@raines1220

2 months

The first second in FSD v12.4.3, my butt already feels something different. Will test more tomorrow

9

1

53

Raines

@raines1220

4 months

@Tsla99T 像SpaceX、Tesla这样的公司市值上限在哪里，我觉得是很难估算的。一般认为公司成长速率会越来越慢，而未来的现金流都需要被折现。那么长期来看，公司未来所有的现金流就会收敛到一个确定的数值。

6

2

52

Raines

@raines1220

27 days

One of the risks of FSD that I previously predicted—and was later confirmed by Elon—is the lack of diversified data for validation. As FSD improves, the availability of useful data (e.g., interventions) decreases, making it increasingly difficult to determine whether a new model

13

2

53

Raines

@raines1220

4 months

My gut feeling tells me that FSD v12.4.1 will roll out this weekend, but the NN I trained predicts that FSD 12.4.1 will be released on May 30th (± 2.2 days). Should I trust myself or the NN? 🤯

13

0

50

Raines

@raines1220

16 days

I realize that I’m not making excuses. Driving 15 miles to buy a watermelon? That’s my real desire. It may be a small one, but it’s genuine. Previously, I had so many minor errands that required driving, and I would refrain from doing them because they didn’t seem worth the

2

4

51

Raines

@raines1220

18 days

@WholeMarsBlog People probably don't understand that when they got FSD v12.5.1.3 now, it was literally much better than my FSD v12.5.13 when I first got it. The slow rollout is not only about validation but also about conducting A/B experiments and finding the best NN from all the candidates in

4

1

52

Raines

@raines1220

2 months

@Tsla99T 特斯拉目前应该没有理论上“光子进，操控”出的理论完全“端到端”，因为大屏上的路径规划依然与现实中v12的行为是一致的。如果真的成为一个大黑箱，那么这种“中间数据”不可能这么“可解释、可监督”。

11

1

51

Raines

@raines1220

1 month

After more driving on v12.5.1, I felt the model is less safe but smoother. Here’s my guess: v12.5, as a new model, included many heuristics to maximize the safety of beta testers with the “advanced update” enabled, such as the creep limit and subtle brakes in traffic control.

2

5

50

Raines

@raines1220

2 months

But eventually, I think they will make it fully end-to-end because it helps the real-world AI understand its own body better. We still see v12.5 making turns that end up riding the line and then fixing themselves later. Humans also make these small mistakes, but AI should do

1

2

46

Raines

@raines1220

1 month

$TSLA Q2 has $622M in restructuring charges, which are one-time. After adjustment, the actual EPS is approximately $0.66 (about $0.14 higher), which, in fact, is a BEAT to the $0.62 EST.

4

1

46

Raines

@raines1220

1 month

@p_inotao 我觉得理想内部在做PoC的时候可能犯了战略性错误。验证双系统不在于确认其可行性，而在于确认其优越性。他们可能只是拿几个简单场景校验了一下，确保双系统可以做到功能完备，然后就匆匆上线了。

4

3

45

Raines

@raines1220

3 months

@aelluswamy That's why engineers follow where Elon goes. By staying with Elon, you're either working on how to make Starship reenter safely and pass through the blackout zone while streaming globally at SpaceX, or starting with 384 KB memory and eventually solving autonomy and deploying to

3

5

43

Raines

@raines1220

5 months

@TroyTeslike its really bad for a data guy to be obviously biased when designing their survey.

3

0

45

Raines

@raines1220

5 months

FSD v12.3 is almost perfect, especially at night, but it performs poorly in intense light conditions such as 'sunrise/sunset', 'clear skies', and 'clear & cloudy'. FSD v12.3.3 attempted to address this issue, significantly improving performance during 'sunrise/sunset' and 'clear

4

3

44

Raines

@raines1220

2 months

I've already adopted GPT-4o-mini in my business, and it is amazingly good. Our product has significantly improved while reducing costs. I have to say, based on how OpenAI achieved so much with such a small model, I have "fine-tuned" my expectations of FSD on HW3. I am starting to

4

3

43

Raines

@raines1220

3 months

EVERYTHING BELOW IS PURE GUESS, READ WITH YOUR OWN DISCRETION. I think FSD v12 is based on something similar to the MoE architecture. Each subtask is an independent neural network, with probably hundreds of them. They probably trained together so that the gating network knows

8

2

41

Raines

@raines1220

1 month

分享一个业余时间做的AI小工具——AI课代表。支持油管和B站的视频内容总结、字幕搜问、知识提问。我平时分析FSD时，就依赖这个工具来快速了解信息。在浏览器插件商店搜索“AI课代表”就可以找到了。

6

2

42

Raines

@raines1220

26 days

I really want v12.5.1.3. I think the jump from v12.5.1.1 to v12.5.1.3 will be similar to the jump from v12.4.3 to v12.5, possibly due to scaling in training data, as the larger model is far from saturated and is hungry for more data now. Note that it’s only being sent to Model Y

6

1

42

Raines

@raines1220

3 months

“but until we see working prototypes …… we can’t estimate earnings or cash flow ests with any degree of confidence and therefore leave them out of TSLA’s valuation - for now.” This is actually the main reason to invest in $TSLA now instead of later. For those who have AI

Gary Black

@garyblack00

3 months

Institutions are changing how they value $TSLA. The vast majority look at TSLA how I do: 1/ EV business (TAM x EV adoption % x TSLA EV share x ASP x auto gross margin); 2/ FSD (TSLA fleet x EV take rate x $/month); 3/ Energy; 4/ Service/Charging. Many of us see 5/ Robotaxi and

85

84

881

3

4

42

Raines

@raines1220

5 months

Layoffs are key to eliminating internal FUD, which worries me the most, so they can achieve a higher win rate in the Robotaxi startup. That's the secret move that WS never understands, which Elon does a lot during his "wartime CEO mode." Thats exactly how he "fired" Lidar, Radar,

3

41

Raines

@raines1220

29 days

@TheSonOfWalkley For every two months of delay, ~300K more HW4-robotaxis that currently support v12.5 will be delivered. Tesla is scaling its robotaxi fleet even before the business starts. Additionally, for every two-month delay, FSD will be at least 2x as good, thus reducing the need for remote

3

4

40

Raines

@raines1220

2 months

I realized that FSD would flash a high beam for 0.5 seconds just to see the speed limit more clearly during night driving. Is this common knowledge? It completely blew my mind.🤯

6

0

39

Raines

@raines1220

2 months

我个人的一个偏见是，现在很多AI公司是在强行蹭“AI”风口的流量。放在几年前他们可能会被称为“大数据公司”，再放到几年前他们可能被叫做“软件外包公司”。想从“软件外包”真正进化到一家人工智能企业，我个人偏见以为：在算法以外，必须拥有数据的所有权，或者拥有超算中心的所有权。在算法，数据，算力这

9

1

41

Raines

@raines1220

1 month

The Cybertruck is now on version 2024.26.4, while other vehicles will soon be upgraded to FSD version 12.5.1.2 (2024.26.5). It seems they are finally on the same 2024.26 branch and just one version away from being merged together. I believe that in FSD version 12.5.1.3 or 12.5.2,

3

40

Raines

@raines1220

2 months

I got FSD v12.4.3. I can’t wait to prove myself wrong.

3

0

39

Raines

@raines1220

25 days

FSD v12.5.1.3 has a parking addiction. Whenever it sees a parking spot, even if it is not the destination, it cannot resist parking itself.

9

0

38

Raines

@raines1220

1 month

Soon, FSD will need to supervise human driving and intervene if necessary. Grok will evaluate human drivers at scale and comment on which human is the best driver based on how many miles a human can drive before FSD intervenes.

4

3

37

Raines

@raines1220

2 months

I'm not gonna tell everyone that I went all-in with leverage on $tsla, but there will be signs.

Lior

@liorsela

2 months

@daltybrewer I’m not gonna tell anyone I own $TSLA but there will be signs

8

4

258

4

1

36

Raines

@raines1220

20 days

One thing I recently realized, which I didn’t fully understand before, is how to interpret FSD intervention data. When we talk about FSD driving 100 miles between interventions, it doesn’t mean that FSD will randomly try to “kill you” every 100 miles. FSD is actually quite

Ed Hill

@aronth5

21 days

@raines1220 For many others there have been very few regressions. One of the great mysteries of FSD is why does the driving experience vary so much for different people?

0

1

4

1

37

Raines

@raines1220

2 months

1. AI is the ongoing industrial revolution. Along with Nvidia "bullying" Moore's law, the US banning AI chips to China ensures that US companies remain leaders. 2. Tesla has the largest fleet generating the most diversified data. The best CEO decided to go "balls to the wall" by

Ayden Miller

@AydenMMiller

2 months

What are 2 reasons that you’re invested in $TSLA? I’ll start… 1. $TSLA is a highly PROFITABLE long term investment… (FSD/Optimus/Energy) 2. Tesla is at the leading edge of AI/FSD which will be the fastest growing industry worldwide…

14

4

47

3

36

Raines

@raines1220

4 months

Based on previous research, I've noticed significant differences in the performance of FSD across various states. This observation highlights the importance of assessing "dataset concentration" when evaluating crucial metrics like "city miles2DE crit." To this end, I decided to

3

6

36

Raines

@raines1220

1 month

I'm >60% sure that v12.5 now recognizes hand gestures.

ΛI DRIVR

@AIDRIVR

1 month

FSD 12.5 threads the needle 😬😬😬

89

179

3K

0

36

Raines

@raines1220

2 months

I'm a huge fan of everything except HW3. Regarding HW3, I'm a FUD. FSD on HW3 cannot think. It is mimicking the behavior of thoughtful thinking. Putting more and more complicated scenarios into training just makes it overfit. It lane dances because it believes more lane changes

12

1

36

Raines

@raines1220

29 days

I always assume AI has feelings and emotions, though they exist in forms we don’t feel or understand. Imagine our creator watching this world of simulation and laughing at us: “I gave them a simple loss function, which is to survive and reproduce, even though everyone eventually

2

1

37

Raines

@raines1220

4 months

Can’t sleep, any suggestions? 👀

10

0

35

Raines

@raines1220

3 months

@elonmusk @SawyerMerritt If I understand correctly, Tesla internally categorizes all causes of interventions. One category is called "unknown," and it happens approximately every 15,000 miles (the average miles driven per year). If they manage to solve all known “bugs,” then it is technically true to

5

2

35

Raines

@raines1220

3 months

@ByeonChansoo “Only when the tide goes out do you discover who's been swimming naked.” —Warren Buffett

1

34

Raines

@raines1220

3 months

One day, people blindly love robots made by TSLA. They are a bit more expensive than their competitors, but people only buy TSLA bots because they are smoother, smarter, and all TSLA robots share the same memory of you, knowing your preferences and lifestyle better than your

TheSonOfWalkley

@TheSonOfWalkley

3 months

BREAKING: WARREN BUFFETT SELLS BYD STOCK, NOW ONLY HOLDS 6.9% 👀 $TSLA Will Warren Buffett buy Tesla stock now?

82

59

1K

4

2

34

Raines

@raines1220

25 days

[2/2] Another disengagement is debatable as to whether it is a critical one or not. What happened is that FSD's view was almost fully blocked by a large truck. When it saw the light change from yellow to red, it had no reaction time. But don't get me wrong, FSD did the right

7

1

35

Raines

@raines1220

19 days

Recently, I've been thinking about the relationship between miles-to-intervention and miles-to-accidents. According to the National Safety Council, motor vehicle collisions between two cars account for 71% of all incidents. If we consider that all accidents are basically two cars

7

2

35

Raines

@raines1220

1 month

理想的双系统技术方案，属于技术人典型的路径依赖型的战略错误。陷入技术验证的泥潭，忘了验证商业逻辑。但凡有点第一性原理的思维，都不会这么鲁莽。

pino

@p_inotao

1 month

@raines1220 内部的决策不好猜测，不过通过公开采访看。理想花了不少时间和资源让英伟达的Orin能够跑起它的2.2B的VLM，且过程做了一轮的性能调优。现状就是勉强能跑了。我们也知道不管LLM还是VLM，幻觉的问题是工程化落地的一个x因素。能不能调到一个可用，我后面的版本再持续观察看看。beta2马上要推了

1

0

3

9

1

34

Raines

@raines1220

3 months

@Tsla99T 我在问界群里问了一样的问题，现在已经被踢出群聊了。

4

0

34

Raines

@raines1220

4 months

I'm sure you already know that North America has ~1.8M customers who might use FSD v12, and half of them have already tried it during the free trial. Did you also know that there are ~1.6M potential FSD customers in China? Previously, their take rate was 0%. Now, FSD will soon be

5

7

35

Raines

@raines1220

1 month

Based on my latest findings, there is no "dot version regression" in FSD. The statistics often misused by $tslq typically use the community tracker to illustrate how FSD "regressed" from v12.3 (29 miles2DE) to v12.3.6 (28 miles2DE) and from v12.5 (115 miles2DE) to v12.5.1 (39

5

1

35

Raines

@raines1220

19 days

The only difference is the distribution of safety events. Human errors caused by distractions, reckless driving, etc., are more evenly distributed. In contrast, FSD is still constrained by its intelligence, making it more fragile in difficult scenarios, leading to a more

4

1

35

Raines

@raines1220

20 days

Elon: How long do you think is needed to make it work on HW3? Team: Well, it is a huge amount of work. We need to rewrite everything to make it no longer uses CPU......I cannot guarantee, but I would say at least one month. Elon (posted): The estimated rollout is 10 days.

3

1

34

Raines

@raines1220

13 days

I love your video. I noticed something very special: a very long trajectory produced by the 12.5.1.5 path planner. I’ve seen it several times in my car in shadow mode, but it jiggles like crazy, and I thought, “Thank God it’s in shadow mode.” However, this is the first time I’ve

⚡️TechGeek Tesla 🔋⚡️

@jonbbc

13 days

The zero intervention drive across 45 minutes w/ FSD 12.5.1.5 on HW3 in Chicago is live on my channel. I was up all night cranking this one out. 🫠 Couldn’t sleep!! This update is very impressive. Lane selection has been near perfect and so have the lane changes. I’ve only tested

29

28

348

5

1

34

Raines

@raines1220

4 months

When AI surpasses humans, it will be like a train catching up to you. It seems really slow when you see it from afar. But once it catches up, you realize that no matter how hard you try, you'll never be able to catch it again. This has already happened multiple times, such as

3

1

32

Raines

@raines1220

4 months

I analyzed Tesla's historical votes for the past 6 years and here are some insights: 1. Board Recommendations had an average 76% Win Rate on voting results over the past six years, suggesting that the board is greatly trusted by their shareholders. 2. When votes were approved

3

2

32

Raines

@raines1220

3 months

Believe it or not, regression on lane indecisiveness in FSD v12.4.1 is a good thing and a strong indicator that structural changes have been made. We might get out of a local maximum and quickly descend to a new point where most lane issues could be resolved soon.

2

1

31

Raines

@raines1220

2 months

ok now we have the final evidence that v12.5 has 5x parameters.

Elon Musk

@elonmusk

2 months

@WholeMarsBlog We are focusing on just Model Y with HW4 for the initial release. Make sure that works well, then broaden. This has the 5X increase in parameters. HW3 would run the same parameter count, but requires extra work to optimize the code.

334

389

4K

2

1

31

Raines

@raines1220

23 days

FSD TAKE RATE AT 39%-58%. As of Q4 2022 (December 31, 2022), the FSD fleet was approximately 400K, with the North America HW3 fleet at 1.34M. During this period, the FSD fleet drove roughly 11 million miles per month, averaging about 30 miles per car monthly. Based on my

3

1

32

Raines

@raines1220

2 months

The driver's foot was on the accelerator pedal, forcing the FSD to accelerate from 35 to 38 and run the stop sign. It was at the last moment that the driver moved his foot to the brake.

Dan O'Dowd

@RealDanODowd

2 months

This angle shows how it looked from inside the car when FSD blew past a stop sign at 35mph, and doesn't have the audio of car FSD nearly T-boned honking at us edited out:

36

15

89

4

1

31

Raines

@raines1220

1 month

"Believe me, there’s nothing better than buying from someone who has to sell regardless of price during a crash." —— Howard Marks

0

1

29

Raines

@raines1220

4 months

NEW RESULT: Assuming Elon isn't lying, the closest metric we have that compares to 'miles-per-intervention' is 'City miles to DE.' These are not the same thing, but let's assume they are for now. v12.4 would be 5x-10x better than v12.3.6, which has 15 City miles to DE. This would

4

6

30

Raines

@raines1220

3 months

@RudyTechEvo My BS detector alarms like crazy whenever a company claims to have built super-advanced real-world AI without a known source of their training dataset.

2

0

30

Raines

@raines1220

4 months

So exponential growth does not necessarily require a "minor" release like v12.4. To be honest, this is huge... @a_meta4 @Tsla99T @onenewbite @ValueAnalyst1

1

3

29

Raines

@raines1220

2 months

@TheSonOfWalkley let me guess, you are literally SHAKING now and only 41% down on your all-in @ $414 position… btw, LFG!

9

0

28

Raines

@raines1220

1 month

Fuck it. I went out. The first thing I notice is that the close-to-curb problem seems solved.

1

0

29

Raines

@raines1220

4 months

@elonmusk @WholeMarsBlog A 5x-10x improvement from v12.3 would likely guarantee that the Robotaxi business won't lose money, according to my model. Any additional improvements starting in v12.5 would just increase the business's margin.

7

0

28