When unsupervised FSD is released the value of the US Tesla fleet will go from ~$100B's to ~$T's. I'm going to leave this prediction here and see how it ages. I feel pretty good about it.
To ChatGPT:
write a short essay about how news used to be controlled by concentrated media organizations but has now become more democratized
Response:
For decades, traditional media organizations have controlled the distribution and consumption of news, limiting the…
I just had a 2hr FSD drive on winding mountain roads that was transcendent. It was perfect. Not good: PERFECT. Smooth, controlled, assertive, COMFORTABLE. Better than human. Oh man this is really happening 😁😁😁. Bravo AP Team. Bravo
@elonmusk
Real time access to X posts really is a killer feature for Grok. Conversational and tailored access to timely current event summaries is awesome. This alone is worth the subscription.
V12 is going to be over 100x reduction in interventions compared to V11. This is not an incremental upgrade, it's a leap forward.
Thanks
@Rebellionaire3
for giving me a chance to experience V12 path planning and control. It's impressively better than V11 and I predict that most…
The body dynamics are so good I’m having a hard time believing this video is real. This is the neural network based body control I’ve been waiting to see. Robot movement should be efficient, graceful, even beautiful. Impressive progress.
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 🧘)
→
On the topic of Tesla's Safety Report: I keep hearing assertions that these numbers are not representative of actual safety because AP is only used on roads with few accidents. This seems to be FUD. 🧵 /1
Robotaxi Teslas will drive 5x the miles of personal cars. This is:
5x displaced ICE miles
5x+ value to users
5x++ profit to investors
Environment, customers, and investors are perfectly aligned here.
Tesla can keep making personal cars but every effort should be made to…
A nicely constructed and explained napkin-math valuation for Tesla's FSD in the referenced tweet below.
You might prefer different assumptions but conservative numbers give FSD a valuation that is 10x the value of just selling cars. The assumptions in the model below are modest.…
Let's have some fun "valuing" Tesla FSD + Autonomy...
... Introducing the "CERN" Model for Tesla - a Conceived Estimate of Robotaxi Network.
Let's look at just the potential value of Full Self Driving "FSD" software + Autonomy. (Note: I say "potential" - as this isn't a company…
Today Tesla release a whitepaper describing new numerical formats being used in Dojo. These novel numerical formats are created with the needs of neural network training in mind. A brief explanation follows: 1/6
This is wrong. The biggest factor *by far* is making the tech work well. When it does the benefits are too big and too visible to be masked or argued away. How many places prohibit fire, plumbing, or electricity?
The biggest factor in the timing of Robotaxi is out of $TSLA's hands, state approvals.
Crazy to think the hardest part of autonomy is getting past lawmakers.
I'm frequently asked why is a car company building a supercomputer for training and this question fundamentally misunderstands the nature of Tesla. At its heart Tesla is a hardcore technology company. All across the company people are working hard in science and engineering to…
Improving FSD
I've been asked a few times about how and whether FSD can continue to get better without improvements to the hardware. This question is often asked in the context of a misunderstanding of how neural network optimization and improvement generally works, so I'm going…
4 hours on 11.4.7 today. Heavy LA fwy traffic and downtown surface streets, coastal highway and winding mountain roads. No interventions, no wonky behavior, no anxiety. FSD is getting so good.
I spent a 5-month internship in Tesla's Autopilot team and it was such an amazing experience. I worked with super dedicated and talented people while solving one of the hardest robotics and AI challenges. The team is
@CVPR
and is hiring (
@philduan
)!
@WholeMarsBlog
Not quite betting the company, but going balls to the wall for autonomy is a blindingly obvious move.
Everything else is like variations on a horse carriage.
Have to see how consistent this is but looks like confidence in predicting other cars is up. High confidence lets FSD execute smoothly without excessive margin.
Here is a highlight clip from my first drive on
#FSDBeta
v11.4.2. This drive was really very good. Full length will be posted on YouTube and in my $1
@Twitter
subscription channel after rendering is complete.
@elonmusk
@aelluswamy
An interesting aspect of FSD (not FSDbeta!) V12 is the applicability of scaling laws. This should let Tesla predict the performance of an FSD implementation ahead of time. That’s important.
Scaling laws in neural networks are stable relationships between the performance of an NN…
I don’t believe in FSD because I’m bullish. I’m bullish because I believe in FSD. I was looking at DL for self driving before I knew about TSLA: the tech led me to the stock. I wasn’t interested in EV stocks. I worked on EVs in the 90s and I knew it was a tough business.
AI and robotics will be a gravitational singularity of a technology transition. Their influence will warp almost all economic activity, sucking in huge chunks of the economy. The ripples in reality flowing out of them will spawn new worlds. Tesla is one of very, very few…
The best feedback to the AP team is an intervention. When FSD is being less than excellent override its behavior with the accelerator, brake, or steering as appropriate. The intervention will be noted and logged and the captured data can be used to train future versions.
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).
Interesting choice of words. Maybe accidental but it suggests they have settled on the architecture and expect that with enough training the current NN will be a release candidate. That's promising.
Just before the start of this clip there was an interesting maneuver by V12 (which we are talking about at the start of the clip) which is illustrative of the difference between how V12 works relative to V11.
Our car was waiting to make a left turn at a light. It was in a left…
Tesla FSD v12 meanders through a parking lot with James Douma (
@jamesdouma
).
This is the clip Bradford referenced in the video with Farzad from earlier today.
“The reality of doing the right thing is more important than the perception of doing the right thing.” -Elon Musk talking about criticism of Tesla’s FSD efforts
Elon: (In response to question implying need for $25k car) "It's apparent from the questions that the gravity of FSD is not fully appreciated... If the cost of our cars does not change *at all* we would still sell as many as we could possibly make."
Of the pieces needed to get to robotaxi the fleet is the most important part. No significant component will be left out because it’s an enormously valuable and difficult to substitute resource.
@sobieski902
Pretty much.
It has been staggeringly difficult to make generalized self-driving work, requiring all that you describe above and more.
The investment in training compute, gigantic data pipelines and vast video storage will be well over $10B cumulatively this year.
But that…
The real robotaxi trolley problem: act to automate driving ASAP and be responsible if someone is harmed by your imperfect system or wait for a perfect system knowing that each day of delay costs thousands of lives.
Solar is the fastest growing energy source ever. This trend is durable and needs to be considered separately from other renewables when forecasting future growth.
Consider Wind+Solar: If we treat them as a group and extrapolate their growth we get the following graph:
On the New Year’s Eve, everybody’s into making predictions
Many come up with futuristic hypotheses, as if competing to single out the wildest, and even the most absurd ones.
Here’s our humble contribution.
What can happen in 2023:
Tesla doesn’t buy enterprise software, they write their own. It’s tailored to their needs and is both better and cheaper than what they could buy. Plus they get control/flexibility/independence. Dojo is similar. It’s good economics. It’s also freedom.
I love this video about dyno technicians. Of course the point of the video is recruiting but my take-away is in a different area. If you are trying to develop hardware that is just fundamentally better then you need crazy good instrumentation to help you understand it and you are…
Driver w/o AP is not improving (red line). So if driver+AP (blue) is improving then AP must be improving. And driver+AP is twice as good as it was 3 years ago.
We aren't going to have nuclear powered AI datacenters. We're going to have solar+battery powered AI datacenters running flexible loads. They'll be built in sunny places with cheap land and friendly regulations. No grid or utilities required if Starlink is used for network…
"08/08 $20k car, they’re going to make $40k or $50k a car a year and if they have 1M cars, 2M cars … they’re going to make 100's of $B a year on that business…" - Ron Baron of Baron Cap
Same as my numbers. Now - what is the present value of $100B/yr of FCF?
@MatchasmMatt
At today’s prices $30 billion would get you 20 GW of solar and 10,000 Tesla megapack 2 XL.
This would basically guarantee you baseload characteristics that are not inferior to the 2.2 GW nameplate nuke while also replacing a large number of peaking plants and average daily…
These changes will bring Optimus’ hands closer to human capability. Matching human manual dexterity is important to the goal of substituting human labor.
Driving with Tesla Autopilot has a large and growing safety advantage over driving without it. Autopilot protects its users and the public as it continues to improve. Unaided humans are not getting safer. If anything they seem to be getting worse.
@MatchasmMatt
@SawyerMerritt
@Tesla
Just eyeballing the numbers it looks like this is approximately a reversal of the large price raises we got during the recent EV demand spike. So margins are going back to their original intended levels (plus COGS improvement since).
Talk is cheap. Words are often mistaken and sometimes they lie. Words are used lightly because they move people but are rarely held to account.
Words are nothing. Results are everything.
Going to the moon is hard. Really hard. But it’s much, much easier than Mars. If your goal is Mars the moon is just a minor waypoint. Similarly general purpose humanoid robots are dramatically harder than robotaxis. Tesla’s odds for robotaxi success just spiked up.
On yesterday's chat with Hans and Farzad I mentioned some stats about rideshare demand. I'll share a few details in a thread here.
This is napkin math but it's based on real demand data and I think it illustrates the business potential.
Once LLM’s could finally write good sentences we immediately proceeded to Kobayashi Maru them by judging them on their human values.
FSD V12 getting hit with the no-win-scenario is one kind of ChatGPT moment.
I had a great chat with
@herbertong
yesterday. We discussed what’s different about V12 and why it represents an important transition in the development of FSD.
Much anticipated, James Douma’s reaction to Elon’s livestream of the FSD V12 drive is finally here!
James
@jamesdouma
has such a great way to explain what’s the breakthrough and what’s not.
A must watch!
👉
The highway stack on V12 is a substantial improvement over the V11 highway stack. Unlike the city stack it's not end-to-end and it isn't a shocking improvement over its predecessor, but it's a solid update that resolves a lot of issues.
This morning I tried four different CCS chargers at two different locations and none of them worked. I struggle to understand what is so complicated about metering electricity. Is Tesla really so extraordinary? Or is everyone else grossly incompetent?
Yes. We don’t know exactly how, but it’s going to be huge. Stay flexible and pay direct attention. Don’t get your opinions from a middleman - they all have agendas.
In a situation like this, there are two general pieces of advice one can give: keep your options open, and pay close attention to what's happening. Then if you can't predict the changes AI will cause, you'll at least be able to react quickly to them.
This was an absolutely brain melting conversation with
@jamesdouma
!
He is such a wealth of knowledge, and he explains his insights in such an easy to understand way.
It was a pleasure, as always!
And shout out to
@farzyness
for doing an incredible job as the host.
Tesla is making a hugely ambitious leap with their upcoming Tesla FSD beta v12 which will use an end-to-end neural network approach for full self driving.
Spoke with
@jamesdouma
to discuss this game-changing move.
(video released first on X and then on YouTube in a few hours)
I love that sound isn’t an afterthought. Doing it well is hard and doing it cost effectively takes clever engineering. Human hearing captures amazing nuance and sound contributes much to our lives. More evidence that Tesla is a product-first, engineering centric company.
120dB+ for a kick drum you can feel in your stomach 🔥
Developed entirely in-house, our audio system provides an excellent listening experience for music, gaming or video content, no matter which seat you’re in
I'm constantly being re-impressed by Rob's ability to made complicated things simple and then deliver them in a no-nonsense, just the facts, minimal speculation tone that is nonetheless enjoyable. We're lucky to have you - thanks
@TeslaPodcast
.
Project Lyra explores the possibility of sending a spacecraft to chase 'Oumuamua using an Earth gravity assist to reach Jupiter, which robs the craft of almost all its speed, causing it to drop to the Sun for an Oberth maneuver.
Thanks
@hibberdadam994
for the trajectory data!
A reminder that there is no one-and-done version of FSD coming. The development pipeline is deep and improvements will continue to be added incrementally for many years to come.
@garyblack00
I do. Elon does. If you know the tech and can independently estimate future capabilities, development, and the commercialization path you can get to 4x the auto business pretty easily even with rapid growth in autos. If your analysis is limited to current financials and what…
Two hours is never enought time for a chat with Dave. We ran out of time before AI armageddon but managed to get through some Investor Day, FSD 11, and GPT-4.
This document enables third parties to consider these formats for forthcoming hardware and software designs so as to enable future Dojo compatibility as well as to share the concept to designers considering how to overcome similar obstacles in the systems they create. /end
@RadarMoron
@elonmusk
@JeffTutorials
@karpathy
You’re right. I didn’t take that earlier comment to mean that they were actually doing the GP part of GPTs, just that it was a transformer. But now that I think about it it would make sense to use GP for a lane prediction transformer.
BTW - not calling out Gary here. We're all bringing different things to this party - his perspective is super valuable and I'm happy that he takes the time to post it. So few people in his position share their honest thoughts with us.
Two models illustrate how important autonomy is to the EV transition. Autonomous EVs displace multiple times more fuel consumption than ordinary EVs, which accelerates the transition significantly. This is a big part of why FSD is central to Tesla’s mission.
2/ FUMES with Autonomy – a rapid decline:
2022 – ICE vehicles drove 3.6 trillion miles and consumed 180 billion gallons of fuel. There are 145,000 gas stations each selling 1.24 million gallons per year.
2025 – EV adoption rises to 16%. ICE vehicles drive 3.5 trillion miles and…
This is Tesla's patent on the Autolabeler concept. It covers any system for extracting neural network training data from time sequence data (like video). I'm impressed that they got such a broad award.
Eventually all surgery will be done by specialized robots. But someone will be the first to develop highly automated semi-autonomous robot surgeons- and that could well be Neuralink.
This is really cool. Tesla examines 6B miles of data and extracts behaviors that correlate with accident likelihood within some window, then you a score for *your* driving. What a great feature.
People misunderstand the value of a large fleet gathering training data. It's not the raw size of the data you collect that matters, it's the size of the set of available data you have that you can selectively incorporate into your training dataset. /1
@j_brorsson
@jamesdouma
@dmitriyTM3
@WholeMarsBlog
@elonmusk
@lexfridman
Green has banged on about how the data gathered from Tesla’s fleet isn’t large enough to be as significant a training set as claimed.
Could the discrepancy photon count data allows for more efficient compression? And that’s why Green isn’t seeing the expected network traffic?
@WholeMarsBlog
@elonmusk
It’s a testament to how pathetic humanoid robot efforts have been to date that, once Tesla builds 1000 of these V2 humanoids they’ll be the leader. That’ll probably happen in 6 months.
Playing with LLMs today - both cloud hosted super-LLMs (GPT4 and Opus) and open source models running on my laptop.
It's so crazy. This was impossible. But here we are. My laptop is a polymath polyglot with eidetic memory and encyclopedic knowledge. The cloud is even more.
No reasonable division gets to 1/3 as many accidents for AP versus non-AP relevant roads. This whole "safety report data not representative" narrative is FUD. I mostly ignore FUD but I just got tired of hearing this one. /end
@DrKnowItAll16
In my case it was uptime. There’s only one cable provider at my address and it’s frequently down for hours at a time. Peak speeds were better on cable but Starlink is fast enough and it’s never down.
The response time of open source ML developers is next level. Interesting model with a novel architecture dropped just before the weekend and we're already seeing performance optimized implementations available to everyone. Go Team!
Adding support for the new Mixtral models
Runs on CPU, CUDA and Metal with quantization support and partial GPU offloading.
Very interesting architecture to play with!
CFP8 and CFP16 address many of the shortcomings of existing formats and allow for 2x to 4x increase in performance and capacity of training hardware, with concomitant improvements in hardware size, cost, complexity, and efficiency. 3/6