Machine
@M_IsForMachine
Followers
2,809
Following
96
Media
148
Statuses
1,844
Explore trending content on Musk Viewer
FEMA
• 1498003 Tweets
نابولي
• 642248 Tweets
2NE1
• 331219 Tweets
Joker
• 172258 Tweets
天使の日
• 132117 Tweets
Grammys
• 100633 Tweets
Mancuso
• 99639 Tweets
Song of the Year
• 63697 Tweets
名探偵ピカチュウ
• 56792 Tweets
Record of the Year
• 55487 Tweets
Best Record
• 54059 Tweets
Ülke
• 52989 Tweets
Pogba
• 46942 Tweets
AFFAIR EP6
• 45005 Tweets
Ocho
• 37543 Tweets
Bielsa
• 37442 Tweets
Trump Bible
• 37003 Tweets
Uyuşturucu
• 31611 Tweets
HAPPY 2nd BIRTHDAY SMILEY
• 25314 Tweets
Celtics
• 16699 Tweets
Jamie Dimon
• 13478 Tweets
菅野メジャー
• 12502 Tweets
Semih Çelik
• 11979 Tweets
#اوقفوا_الاباده
• 10121 Tweets
Chart of
@Tesla
FSD progress vs a realistic approximate goal to beat human drivers.
(Note the latest data point needs a lot more data to get a stable estimate)
80
16
248
The man who constantly warns about the dangers of AI is now disseminating fake AI videos for political reasons.
Keep logic in your brain please.
55
26
200
This is happening in a lot of Tesla folks heads.
$TSLA $NVDA
27
16
173
This afternoon
@WholeMarsBlog
mashed the accelerator of his new Model 3 from a dead stop when I wasn't paying attention.
Nailed my head against the headrest pretty good. Almost as hard as he nailed my mom.
40
8
166
This software is legit smooth.
Giving
@M_IsForMachine
his first ride on Tesla Full Self-Driving Beta 12.3, with
@a_meta4
in the back seat
Watch in 4K:
Highlights
0:00 — Changing lanes across entire road
0:45 — Car passing on tight street
1:05 — Confusing intersection
1:20 — Car
69
136
819
12
6
156
My current prediction of Tesla FSD Timelines:
November 2024 - FSD on HW4 reaches 1,000 miles per critical disengagement on current model architecture w/ additional data ingestion and longer training. HW3 slightly worse performance.
January 2025 - Tesla releases TeslaRide app to
58
17
143
If you need to exaggerate to communicate your point, it means your position is weak.
Try again.
Kamala is quite literally a communist.
She wants not merely equal opportunity, but equal outcomes.
67K
150K
624K
46
9
132
What a petty douchebag.
And missing the big picture.
31
9
86
@Junheng_Li
@BradMunchen
1) Tesla partnered with Baidu for maps
2) Tesla doesn’t need as much compute in China. They can import U.S trained FSD model and “fine tune” with Chinese data locally, which is less computationally expensive.
3) Just speculation
11
3
115
so
@ValueAnalyst1
remember when I said I "felt" 12.3 reduced critical disengagements by 2/3rds aka 67%?
Looks like avg miles / disengagement (~ 100 before) went up to 264 (so far), aka a 63% reduction in disengagements.
Who's your daddy?
12
5
114
This account is a big turd.
The difference between my CEO and other CEOs on July 4th, one is building the next generation Tesla Gigafactory supercomputer clusters, while the other is surfing.
Not all billionaires are created equal.
401
286
3K
18
8
112
(1 / 2) Have
@Tesla
car prices finally bottomed?
It finally looks like a yes.
After seeing large % declines over past 21 months, used Tesla list prices have flattened the past 2 months.
In fact, Y and X have increased slightly.
First time since June-July 2022
16
5
95
First time I’ve ever seen him quoting these FSD stats!
@alex_avoigt
We have a clear path to doubling average miles between interventions in August
227
418
4K
16
3
88
Prepare for low Q1 deliveries, low EPS, and a share price of $150 in the next few months $TSLA
32
10
76
If you think
@baglino
and
@rohanspatel
resigning the same time as mass layoffs announced is a coincidence…
I would suggest sticking with index funds.
15
10
73
@TSLAFanMtl
@Tesla
Is this a joke, I made it log scale to make the goal look closer.
Do you want non-log scale?
9
0
74
This will finally meet the specs
@WholeMarsBlog
requires in a girlfriend.
8
1
73
For better or worse, this basically guarantees a Trump election victory.
I fully endorse President Trump and hope for his rapid recovery
90K
371K
2M
12
2
66
Tacit admission that HW3 won’t be able to do robotaxis as I’ve long predicted.
Give up
@a_meta4
@ValueAnalyst1
@TeslaTrip
@TeslaTrip
@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
401
404
4K
30
9
63
Wow. The head moderator (and long time Wall Street investor) of
@TeslaMotorsClub
investment forum crushing
@elonmusk
for his intentional vagueness in terms of FSD progress.
Fascinating to see some people (and some employees I know) fed up. Surprising
18
11
62
Who is going to tell Tesla fanbois about energy absorbing crumple zones?
Cybertruck rear ended by ICE pickup
ULTRA HARD STAINLESS!
108
195
2K
22
5
58
My bullshit detector is ringing hard.
16
5
55
While I disagree with many of
@VP
Kamala Harris' policies, nothing is more potentially destructive to the stability and trust in the American economy than
@realDonaldTrump
continuing to sew doubt in the validity of the 2020 Election voting.
19
3
57
@a_meta4
@jamesdouma
@BlackRock
@elonmusk
@aelluswamy
If they are actually really confident in their process, they should be willing to share how they will fit such a high fidelity model on limited compute.
11
2
58
@a_meta4
@jamesdouma
@BlackRock
@elonmusk
needs to give a real plan. "Trust me bro" isn't going to work. No
@aelluswamy
will have to say more than "it's the beginning of the end" to instill confidence.
2
4
58
Apparently in an email to
@Tesla
employees,
@elonmusk
wants employees to put in 80 hours a week.
Does 80 hours on
@ValueAnalyst1
Spaces count?
19
2
56
Bradford might be the most realistic long term $TSLA investor. Been in the stock much longer than most but hasn't turned into a social media hype machine unlike others.
$TSLA Gamma Squeeze? ❌
$TSLA Short Squeeze? ❌
$TSLA going to $500 in a hurry? ❌
No, no, and no. imo 🤏🧵
Sentiment around $TSLA was very low going into the vote. It was so bad, some claimed the stock wouldn't go up on a For vote. That was the fear talking and we said so. 1/
54
33
591
7
1
54
I never pull out
@ValueAnalyst1
16
2
53
13
2
49
Just had a female doctor poke and prod my testicles while I made multiple jokes.
There's a first time for everything.
19
1
51
So 12.5 probably 10-20x improvement. Nice
@WholeMarsBlog
Yeah, seeing it everywhere.
Btw, 12.4 goes to internal release this weekend and limited external beta next week. Roughly 5X to 10X improvement in miles per intervention vs 12.3.
12.5 will be out in late June. Will also see a major improvement in mpi and is single stack – no
504
828
6K
11
1
48
Tesla's FSD? We don't know accident rates, but we have insight into critical disengagements. According to , the best version of FSD V12 can go ~ 300 miles / critical disengagement.
6
3
46
People smarter than me like
@a_meta4
and
@jamesdouma
feel free to chime if you see glaring errors in these assumptions. I'm not acting like my thought process is foolproof.
3
3
44
@TeslaTrip
I'm glad we have semi-retired people with the time to go drive the same route 10 times a day 😉
3
0
43
After closing out my $TSLA puts at $160, I’m now scaling in buying some more.
Long term bullish, but short term people are buying into FSD hype not thinking about bad earnings.
Making money still matters
20
2
44
I’m going to win that 99%+ drives without critical disengagement by end of September
@ValueAnalyst1
IT IS DONE.
My friend's Tesla on v12.5 just drove us around flawlessly for 2 hours. 50 miles on city streets.
It drove _better_ than I can, smoother on acceleration, no flinches or hesitations.
No disengagements and I only did one accelerator press to have it go faster.
165
474
3K
5
1
43
The problem with social media is people get walled off and downward spiral from reality. Twitter is no better, and the owner has turned into a complete dirtbag of personality.
This is not about political views.
21
3
43
If you have a XY chromosome pair, you probably shouldn't be competing against XX chromosomes females.
6
0
41
I posted this when $TSLA was at $202.
It hit $153.75 earlier today. Close enough. Closed my puts for now.
Prepare for low Q1 deliveries, low EPS, and a share price of $150 in the next few months $TSLA
32
10
76
12
2
41
Concerning
@elonmusk
Positive Opinion of Elon Musk:
Negative: 45%
Positive: 34%
NBC / Sept 17, 2024 / n=1000
667
605
12K
19
2
54
A critical disengagement happens when the car is about to seriously screw something up. That could be as bad as get into an accident, but also more "minor" hiccups like hitting a pothole or a curb. Do those get counted as accidents in Waymo's figure? I don't know for sure
3
1
40
Let me just say Neurology MDs from a prestigious university in Palo Alto say Biden shows clear affects of Parkinson’s.
Mouth agape, stare into distance, rigidity.
Americans want a president
Not a dementia patient
17
7
156
14
6
38
Not joking, probably
@karpathy
's goal of "Operation Vacation" now running the FSD training pipeline.
2
1
38
Still, lets assume only 20% of critical disengagements are to avoid what counts as an accident. That means the target critical disenagement rate is not 160k miles per, but 32k miles per.
In other words: FSD needs to improve 100x to match performanance of the human.
7
2
37
I have an old account with 40 followers. I just noticed....
How in the world is
@BarackObama
one of them? 🤔
13
2
37
@NateSilver538
The worrying fact isn't how Joe Biden is today, but what will he be like in 2-3 years? It's only going to get worse.
10
1
37
By virtue of using the FSD software as a curation filter, Tesla can increase data that is highly focused, and I hypothesize can increase error-reducing potency of that data per epoch of training vs feeding in all data.
3
2
35
The plots on the left show error rate (log scale) vs amount of compute used. First, notice the overall trend: To get an order of magnitude improvement in error rate, you need at least an order of magnitude increase in compute (among other things).
5
2
36
If true, stock headed to $130.
You simply can't expect earnings growth to be so bad (let alone negative) and think the stock can have a higher forward PE ratio than other tech companies. Market simply does not care enough about FSD (yet) to overcome that simple fact.
I've updated my $TSLA EPS model based on the Q1 P&D report. I'm assuming the decreased utilization is a fairly significant drag on manufacturing margins, and I'm estimating $0.42 in Adj. EPS. This would be a 50% decline in YoY earnings for $TSLA.
I don't think there are
49
18
298
17
3
35
(1 / 4)
Current State of the 2024 Presidential Election:
Based on current data,
@realDonaldTrump
has 268 Electoral Votes in the bag.
A candidate needs 270 to win.
All these states in Red are showing clear, consistent voting margins in Trump's favor.
7
5
34
Is this a limiting factor? I don't think so. If Tesla plans on increasing compute capacity 10x as they have planned to do, then that combined with running the training job for 10x as long should get you there. Sure it may take 10 weeks to run the training job, but still feasible
5
2
34
And as you reduce error rate by the first 10x, compute requirements to get the next 10x increase further. So, aside from other requirements, in order to reduce error rate by 100x, 𝐲𝐨𝐮 𝐰𝐢𝐥𝐥 𝐥𝐢𝐤𝐞𝐥𝐲 𝐧𝐞𝐞𝐝 𝐜𝐨𝐦𝐩𝐮𝐭𝐞 𝐭𝐨 𝐢𝐧𝐜𝐫𝐞𝐚𝐬𝐞 𝐚𝐭 𝐥𝐞𝐚𝐬𝐭 𝟏𝟎𝟎𝐱
5
3
35
But what about inference compute? This is the biggie. How much can we compress the model to use on Tesla's HW3 or HW4 inference chips? This requires model distillation which has shown the ability to signficantly reduce model size, usually at some cost to accuracy.
2
2
34
Chinese won’t be willing to pay as much as Americans. American take rate is likely lower than 50% at $99.
Maybe 30% take rate at $50/mo
~ $300 million in annual revenue or $0.1 in EPS.
See how making realistic assumptions matters?
$TSLA has sold a total of 1.7 million vehicles on China. If half are eligible for FSD at $99/month, that would be ~$84M in monthly revenue and ~$1B annual revenue.
60% margins would mean over $600M profits annally for $TSLA. Obviously, this will quickly grow exponentially.
Add
136
50
742
16
0
31
So, what does it take to train a deep neural network to improve 100x in less than a few years?
We can examine neural scaling laws to better understand how compute, data, and model size relate to increased accuracy.
2
1
33
Take this paper from
@GoogleBrain
on scaling vision transformers to improve predictive accuracy on the ImageNet benchmark.
1
2
34
2 attempts to end
@realDonaldTrump
's life.
And no one is even trying to assassinate
@elonmusk
?🤔
BREAKING: The man suspected in the apparent assassination attempt of Donald Trump has been charged with federal gun crimes.
Follow live updates:
161
172
426
12
2
33
Blocked by
@alojoh
presumably because I corrected him that there isn’t going to be a big March wave in Tesla Europe deliveries as in past years.
People should really respect the search for truth more and less rooting for their “side”.
14
1
32
@TSLAFanMtl
I do believe Raines has the right concept to help control for intersubject variability in reporting. Not sure there are enough datapoints to be sure, but this approach makes sense. Look only at the same people reporting on each version.
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
2
0
32
The dude doping on testosterone injections is gonna tell everyone how to lead a healthy, all-natural lifestyle.
Sure.
What would Kennedy do under a Trump administration? He would be tasked with ENDING the chronic disease epidemic.
Specifically, he said he's going to FIX the corrupt health agencies under a Trump administration.
“These agencies—the FDA, USDA, and CDC—all of them, are controlled
168
2K
14K
24
2
32
This isn’t liberal vs conservative, it’s science vs stupidity.
The abandonment of using science to make informed decisions will lead us down disastrous paths
5
2
30
@TSLAFanMtl
@Tesla
I didn't want to troll with this. Plus I do believe log scale is more appropriate because showing the relative progress (each order of magnitude improvement is a horizontal line).
6
0
31
So, will Tesla need 1 or 2 orders of magnitude increases in data? I don't think it will need 100x, it will be less. Why? These studies focus on increasing data in aggregate, not focused on the long-tail "poor performers".
5
1
31
This means, with respect to training, if we want to reduce FSD V12 error rate by 100x, it seems quite likely the model will have to grow ~ 100x in size. Is this feasible for training? I think it actually is, in theory.
3
2
30
What about amount ofIndividual models are represented by a specific marker size (smaller = less data used) and color (blue = smaller # parameters). We see specific models follow a power law relationship with compute. Smaller models taper out at higher error rates.
3
1
30
Yaman’s mom going through clients.
7
0
28
Tesla fans are still coming to grips with Elon increasingly starting to differentiate HW4 from HW3.
Elon has joined me, you will too eventually.
13
2
26
Yaman this random signed up for X a day ago and already has a space with more listeners than yours ever.
Please explain.
@ValueAnalyst1
8
2
26
There are impressive results, for instance some documented in
You can get a 90% reduction in model size, but usually ImageNet accuracy drops say 10% from 90% to 80% or 80% to 70%.
1
1
28
As I have been saying for years.
Data science- adjacent folks like on Reddit / SelfDrivingCars like to think they are experts but they aren’t.
Waymo lacks geographical data diversity and I’d bet their software isn’t robust across the U.S.
Much like Cruise
2
0
27
Lastly, the biggest factor IMO: model size. Model size is the biggest factor in lowering error rate in these studies (and most). You can compute for longer and with more data, but you will still be capped on performance improvement unless you increase model size.
3
1
25
According to this, most of the cathode materials team has been let go. What does
@LimitingThe
think of this?
Elon Musk is throwing his weight around Tesla, comes in like a wrecking ball by
@fredericlambert
15
21
74
12
1
26
@a_meta4
@jamesdouma
But more important is the 𝘂𝗻𝗰𝗲𝗿𝘁𝗮𝗶𝗻𝘁𝘆 is something serious investors will want more information on before they invest in $TSLA on the robotaxis premise. Do you think large institutional investors like
@blackrock
aren't wondering about these exact issues?
4
1
26
Sounds like my takes on scaling laws and FSD was right.
@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
401
404
4K
14
1
25
I don’t get why
@ValueAnalyst1
and
@WholeMarsBlog
keep saying I’m slacking off at my job.
7
0
25
What is "present day"? Is V12 using up most of the compute capacity on HW4? I'm not sure, but I doubt its only using 1/30th of it.
3
2
25